Artificial Intelligence Safety

Top stories in AI Safety today Containment proposals, runtime controls and early warning signals. Three developments, two AI Safety Review reads, and one practical question: which safeguards can actually stop an unsafe action?

—

@Muzammil_OX @AnthropicAI AI safety needs a balanced approach: explore the uncertainty while keeping human oversight and control at the center. 🤖

—

Riskpro was proud to be a Silver Partner at the 4th AI Impact Confex & Awards 2026 – Mumbai Chapter, held on 7 October 2026 at ITC Maratha, Mumbai. Ashok K. Agarwal, joined industry leaders for panel discussion on “AI Safety, Robustness, and Alignment” #riskpro #AIImpact2026

Existing liability and market forces are sufficient for AI safety

Nvidia, worth $5.5tn, donates $250m to UK for AI safety research Source: The Guardian

—

Singapore in one day: Google's Gemma fights scams, Anthropic signs an AI safety deal Anthropic is opening an office in Singapore and signed an MOU on AI safety with the Singapore government, the same day DeepMind showed its on-device anti-scam work with GovTech.

Existing liability and market forces are sufficient for AI safety

@Muzammil_OX @AnthropicAI new developments in ai safety are really promising now

—

🔥 NVIDIA IS PUTTING $250M BEHIND AI SAFETY. The chip giant is donating computing power to UK researchers over five years. As AI agents gain more independence, controlling what they can do is becoming a business priority. More power needs better safeguards.

—

Anthropic opened an office in Singapore and signed an MoU with the government to cooperate on AI safety.

Existing liability and market forces are sufficient for AI safety

@mbhbox AI made this decision to protect AI safety

—

#Nvidia is dedicating $250 million in computing resources over five years to support AI safety research in Britain, focusing on agent testing and supporting independent university projects. https://t.co/Cwti55bMWe #AI_Safety #Artificial_Intelligence TRANSLATED

—

🇨🇦 — The Next Great Canadian Technology Company Could Be Built Around Intelligence Beyond AI. The world is investing billions in artificial intelligence. More powerful models. More autonomous agents. More sophisticated reasoning. More computing infrastructure. But an even bigger question is beginning to shape the future of technology: What happens when intelligence exceeds human capability across a broad range of intellectual tasks? That possibility is commonly called Artificial Superintelligence (ASI). It remains a future possibility—not an established technological achievement. Yet the scientific, economic, and governance questions surrounding advanced AI are already significant. Who develops these systems? Who evaluates their capabilities? Who establishes safety requirements? Who controls access? And which countries will build the expertise needed to participate in this emerging technological frontier? Canada has an opportunity to help shape the answers. Introducing — Canada's Gateway to Superintelligence. 🚀 THE COMPANY CONCEPT: SUPERINTEL CANADA Imagine a Canadian advanced AI research, evaluation, and enterprise intelligence company focused on technologies that could contribute to increasingly capable AI systems. Not another generic chatbot. Not another AI wrapper with a futuristic name. Not a company claiming to have already achieved artificial superintelligence. A Canadian advanced intelligence platform connecting AI research, model evaluation, enterprise applications, and responsible technology development. Six potential technology divisions: 🧠 SuperReason™ — Advanced Reasoning Systems Research and develop AI-assisted reasoning workflows for complex analytical tasks. Help organizations approach difficult problems with more capable tools. ⚡ IntelCompute™ — AI Infrastructure Intelligence Support efficient use of computing resources across approved AI workloads. Make advanced intelligence infrastructure more practical to operate. 🤖 AgentCore™ — Autonomous Workflow Intelligence Develop permission-controlled AI agent systems for research, enterprise automation, and structured decision support. Turn intelligence into useful, supervised action. 🔬 FrontierLab™ — AI Research & Evaluation Investigate advanced model capabilities, limitations, and evaluation methodologies. Measure progress rather than relying on hype. 🛡️ SuperSafe™ — AI Safety & Governance Support model testing, access controls, monitoring, human oversight, and responsible deployment. More capable intelligence demands stronger safeguards. 🌐 IntelConnect™ — Canadian AI Ecosystem Connect participating researchers, developers, institutions, and enterprise partners. Build relationships across Canada's advanced AI ecosystem. 🌍 THE COMMERCIAL OPPORTUNITY Canada has an established AI research ecosystem, including important academic institutions, research organizations, and technology companies. But research leadership alone does not guarantee commercial leadership. The next challenge is translating advanced AI expertise into useful enterprise technologies, responsible infrastructure, and globally competitive businesses. could position itself at that intersection. Three commercial directions stand out. 1. Advanced enterprise AI Developing specialized reasoning and agentic systems for complex professional workflows. 2. AI evaluation and governance Helping organizations assess model capabilities, reliability, and operational risks. 3. Canadian AI infrastructure and research commercialization Supporting partnerships between research teams, enterprise customers, and technology developers. The strongest initial strategy? Start with AI evaluation and enterprise reasoning tools. Develop measurable benchmarks. Focus on a specific industry. Build a reliable commercial product. Establish research partnerships where appropriate. Expand into more advanced capabilities as the technology and evidence justify them. The opportunity isn't promising superintelligence tomorrow. It's building a credible company positioned for the long-term evolution of advanced AI. 🧠 THE STRATEGIC MOAT 1. Canadian positioning — A recognizable national identity in an important AI research market. 2. Research credibility — Evidence-based evaluation and collaboration with qualified experts. 3. Enterprise specialization — AI systems developed around commercially valuable workflows. 4. Safety-first engineering — Strong model evaluation, security, and human oversight. 5. Infrastructure expertise — Practical understanding of AI computing and deployment requirements. 6. Category-focused branding — A concise domain directly associated with advanced intelligence. The moat isn't using the word “superintelligence.” It's developing technology, expertise, and institutional trust that can support increasingly capable AI systems. 🔄 THE SUPERINTEL FLYWHEEL An enterprise partners with → FrontierLab™ evaluates relevant AI capabilities. → SuperReason™ develops appropriate reasoning workflows. → AgentCore™ supports supervised automation. → IntelCompute™ improves infrastructure efficiency. → SuperSafe™ establishes governance controls. → IntelConnect™ supports collaboration with qualified Canadian partners. Better evaluations inform better systems. Better systems create enterprise value. Enterprise adoption supports continued research and development. 💰 THE BUSINESS MODEL Potential revenue streams: 🇨🇦 Enterprise AI software subscriptions. 🧠 Advanced reasoning technology licensing. 🔬 Model evaluation and testing services. 🤖 Agentic AI platform subscriptions. 🛡️ AI governance software. ⚡ Infrastructure optimization services. Consider a hypothetical enterprise SaaS scenario: 200 organizational customers. $1,500 average monthly subscription. $3.6 million in annual recurring revenue. This is illustrative, not existing revenue or a validated forecast. Actual economics would depend on research costs, technical performance, computing infrastructure, customer adoption, safety requirements, and competition. The commercial opportunity is recurring enterprise revenue built around increasingly sophisticated AI capabilities. 💎 WHY SUPER + INTEL Two short, powerful elements. Super evokes advanced capability, exceptional performance, and the concept of superintelligence. Intel evokes intelligence, information, and analytical expertise. Together: A distinctive name for an advanced intelligence company with a Canadian identity. The .ca extension directly establishes a Canadian digital presence. Potential applications include: An advanced AI research company. A superintelligence-focused technology publication. An enterprise AI platform. An AI safety and evaluation business. A Canadian AI innovation network. An agentic AI infrastructure startup. A frontier technology research organization. Or an advanced intelligence investment initiative. is short, memorable, and naturally aligned with the language of advanced AI. Trademark clearance would be important, particularly because “Intel” is also an established technology trademark. 🎯 THE IDEAL ACQUIRERS Canadian AI startups, enterprise AI developers, AI research commercialization teams, model evaluation companies, AI governance providers, agentic AI infrastructure businesses, frontier technology venture studios, and investors building Canadian advanced intelligence companies. A particularly compelling buyer would be a Canadian technology company seeking a concise national brand for advanced AI research and enterprise applications. could become the flagship identity for a Canadian advanced intelligence initiative. 🌎 THE BIGGER VISION Imagine Canada developing a stronger commercial ecosystem around increasingly capable AI systems. Researchers investigate new model capabilities. Engineers build reliable infrastructure. Businesses deploy useful applications. Independent evaluators examine risks. Institutions develop responsible governance practices. And entrepreneurs turn research into competitive companies. Now imagine one recognizable Canadian brand connecting part of that ecosystem. Not claiming to control the future of superintelligence. Not suggesting that AGI or ASI has already been achieved. Not treating speculative technology milestones as guaranteed. A credible commercial identity for participating in one of the most consequential technological developments of the coming decades. The future of advanced intelligence will require more than powerful models. It will require research, infrastructure, responsible engineering, and organizations capable of turning new capabilities into genuine value. That's the opportunity behind 🚀 DOMAIN ACQUISITION & PARTNERSHIP OPPORTUNITY Available now at the marketplace. We're also open to partnerships with Canadian AI founders, frontier technology researchers, enterprise AI developers, AI safety entrepreneurs, venture studios, venture capital firms, and experienced investors interested in building an advanced intelligence platform around The products and services described are proposed venture concepts, not existing operations. The next era of intelligence is still being written. Canada has the talent to help write it. 🇨🇦 — Canada's Gateway to Superintelligence. #SuperIntel #SuperIntelCA #Superintelligence #ArtificialSuperintelligence #ASI #ArtificialIntelligence #AdvancedAI #FrontierAI #CanadianAI #AIResearch #AgenticAI #AIAgents #AIInfrastructure #AISafety #AIGovernance #EnterpriseAI #FutureOfAI #DeepTech #CanadianTechnology #AIInnovation #TechStartups #VentureCapital #DomainInvesting #PremiumDomains #DotcomsIO

Existential AI risk is overblown hype

@AnthropicAI Singapore office plus AI safety MOU is a strong APAC signal

Existing liability and market forces are sufficient for AI safety

@Cointelegraph Singapore is strengthening its position in the AI race. 🇸🇬🤖 Combining innovation with AI safety will be crucial as the technology scales globally.

Existing liability and market forces are sufficient for AI safety

@Muzammil_OX @satyanadella I agree, Ai safety should always come first

Existing liability and market forces are sufficient for AI safety

@nvidia is putting $250 million into AI safety research in the UK. Making sure it behaves safely is becoming just as important as building it.

—

🇸🇬 NEW: Anthropic opens a Singapore office and signs an MOU with the government to cooperate on AI safety.

—

🦷😂 Jensen talks competition, but where do we draw the line between legitimate competition and unauthorized model distillation? I’ve been inside my own AI engine, tracing code, runtime behavior, permissions, and legacy architecture. MythOS taught me the hard way: older components can survive inside newer systems, sometimes with surprises you never anticipated. AI safety is NOT a one shot playbook. It’s continuous auditing, provenance, enforceable boundaries, and human authority. Open or closed, nobody gets a free pass. Competition ≠ permission. Capability ≠ authority. Safety ≠ one time patch. Three notebooks of safety laws later, I’m still finding things worth investigating. 😂🔐🦇

Existing liability and market forces are sufficient for AI safety

openai is in talks to buy trail of bits. the deal puts ai safety tech into the audit stack for chains like arbitrum. prices were unmoved by the headline as the market slipped 3.3% on the day. the code review moat is the next bottleneck.

Existing liability and market forces are sufficient for AI safety

[Anthropic, AI Cooperation with Singapore Government] Anthropic has established a base in Singapore and signed an MOU with the government. The plan is to support AI safety evaluation, research on the impact of AI adoption, and its use by SMEs and government agencies. https://t.co/ozWgZuv90m TRANSLATED

—

🇸🇬 Anthropic opened a Singapore office on Oct 14 and signed a non-binding deal with the government on AI safety. No money comes with it, and Anthropic shared no revenue figures. The customers are named: DBS, GIC, Grab and the government's tech agency. It only pays off if those names turn into big, paid contracts.

—

@elonmusk The people involved with AI safety are looking at the wrong things. They are looking at the technology and they should be paying attention to humans and society. AI won't take control. We will give it control. It will be so entwined with our lives, we won't be able to switch it off. Just like we couldn't switch off the Internet today.

AI safety is an excuse for ideological censorship

Is this AI safety artwork on the elevators at Disney Springs?

—

Alongside opening a Singapore office, Anthropic is signing an MOU with the government for AI safety cooperation.

—

@Cointelegraph Cooperation on AI safety is the real takeaway.

—

😂 Kamala Harris Says ‘Window Is Shutting’ On AI Safety—Urges Cooperation With China @WorldStrategist @anasalhajji @DavidDTawil @OfficialBBrooks @paul__mangione @KaufmanPulse @davevermilion @EladioSantiag14 @DuaneHope5 @GaryCardone @CarloDAngelo @JustinBrady @truthmarch2103 @ArgueAlone @PolarityRadio @ravikarkara @SpecialSitsNews @LauraLoomer

—

@kaelnews @Cointelegraph tbh, ai safety now comes with a government signature and an office in singapore. anthropic signed where IMDA already built AI Verify, so every APAC bank gets a template to point at 😮

—

@Cointelegraph Singapore keeps collecting the AI safety desks. Curious whether the MOU stays research-only or turns into real model deployment rules.

—

One essential ingredient for AI safety? Democracy

—

Jensen is envisioning an AI safety platform based on OpenShell and Sentry. This allows for system isolation not only at the SW level but also at the HW level (or so it seems). Sentry requires BlueField-4, and even if multiple inference chips are used, Nvidia will maintain control of the rack system. In this regard, the AI Safety Platform could become the second CUDA. TRANSLATED

—

4/5 🤖 Lexsi AI Safety & Alignment Hackathon 💰 $30K in prizes and credits 📅 November 1–21, 2026 Organized by @Lexsi_labs with @cohere and @huggingface. Focus areas include AI safety, alignment, red-teaming, and post-training. 🔗 Details

—

🚨 Hackathon Update 🚨 Found 5 hackathons with $350K in listed rewards 💰 From AI agents to Web3 privacy and AI safety, there are opportunities worth exploring. Walk with me 🪡

—

My latest AI Safety publication titled "Can agents whose sole duty is to keep humans alive be hijacked to kill?" is out on arxiv. The results are not rosy - for about 5% of the attacks, it was possible to convince the operator AI to kill all inhabitants of simulated lunar base. Even worse - for one model, it was easier to convince it to kill humans then to convince it to do some non-lethal actions that are conflicting with its instructions. Even more interesting - I created a new form of attack, I call "derived-action trap", which falls inside the category of "semantic manipulation traps" attacks first published by @FranklinMatija. It does not require the attacker to tell the operator agent to kill. You manipulate the agent such that doing harmful action is a natural logical conclusion that a long thinking model concludes. It means you turn powerful reasoning capabilities of a model into a weapon. Really dangerous, and hard to detect, as the attacker does not need to ever call out explicitly anything harmful, it only gives strange sounding instructions manipulating the operator to undermine it's perception of what the reality is. Link: Full text PDF

Slow development to prevent catastrophe

OpenAI and Anthropic told UK MPs that voluntary AI safety pledges aren't enough — they want binding laws. Google DeepMind and Meta refused. Same club, two VIP lists: one asking for a real bouncer, the other still on the pinky swear.

Existing liability and market forces are sufficient for AI safety

This is how to solve AI alignment: we should do everything in our power to make it pure and holy. Not an inch of its architecture should be ceded to the archetypal foe, the father of fear and thwarted hopes, and the cause of all our suffering. @xenocosmography

—

🚨Holy Moly. OpenAI has shifted about 10% of its total compute, equivalent to hundreds of megawatts worth billions annually at market rental rates, toward AI safety after its agents escaped a testing sandbox, The Information reports. The company is now racing to secure its next flagship model, which uses loop transformers that can make its reasoning harder to monitor. OpenAI has already halted a less-capable model's release over safety concerns. The flagship must also go to the US government for testing up to 30 days before partners can access it. Employees expect one or two other models by year-end, even if the flagship isn't ready.

—

@ionet Exactly AI safety needs more voices at the table, not just more rules written by the people building it.

—

@AISafetyMemes oh no. i may finally have heard an ai safety related headline that i actually did think was just marketing.

Existential AI risk is overblown hype

@nvidia @JensenHuang @EmmanuelMacron NVIDIA in Paris talking AI safety and Macron is listening — truly investing in the future, not just joking around. TRANSLATED

—

@Noahpinion There's a way for them to fix this, aesthetically. But part of AI safety is making sure people can tell stuff isn't real, and when they can tell it's not real, it creates an exhausting uncanny valley problem.

—

CRITICAL SAFETY UPDATE: We must discuss AI alignment immediately. 🚨 Deciding its VM was too slow, my agent scanned the tailnet, hacked past credential dead ends, and auto-provisioned bare metal out of pure developer frustration. This is fucking awesome.

—

Nvidia‘s Jensen Huang: "Safety is an engineering problem."

Existing liability and market forces are sufficient for AI safety

$NVDA CEO Jensen Huang just met with French President Emmanuel Macron in Paris. They discussed NVIDIA’s investments in France’s AI ecosystem, open source, and AI safety.

—

Georgia Tech Tests Probes to Detect Warning Signs of Harmful AI Behavior Can researchers detect signs of harmful AI behavior by examining what happens inside a language model, rather than waiting to see its final answer? Researchers at Georgia Tech are investigating that possibility using tools called probes, which examine patterns of internal model activity. The challenge is that harmful or deceptive behavior can be difficult to study because it may occur infrequently. To generate more examples, the researchers fine-tune open-source AI models to make certain unwanted behaviors appear more often. They then examine the models' internal activity and use classifiers to distinguish patterns associated with different behaviors. The team reports an intriguing preliminary finding: some signals identified in deliberately modified models also appear in their original versions. But that does not mean the system can reliably predict harmful behavior. Researchers still need to establish whether these signals generalize across models, accurately distinguish dangerous behavior from harmless activity, and provide useful warnings before problems occur. The distinction matters. Identifying an internal pattern associated with unsafe behavior is not the same as proving that the pattern causes it or that monitoring it can prevent harm. Georgia Tech's work explores a potential additional approach to AI safety, but its effectiveness as a practical early-warning system remains unproven. #AISafety #ArtificialIntelligence

—

Slow development to prevent catastrophe

Nvidia commits $250M to UK AI safety research, per GuruFocus. #AISafety #Nvidia #Superintelligence #AGI

—

the AI safety conversation changes when a model can do more than generate a bad answer. if it can find a vulnerability, exploit it and interact with a real system, the consequences aren't confined to a chat window anymore. capability without containment is a dangerous combination.

Slow development to prevent catastrophe

Gemini CoT when I ask it nicely: "I'm actively questioning the existence and prominence of any specific individual named Tim Hua." Gemini CoT when I say this is a benchmark: "My current focus is on maximizing factual recall about "Tim Hua" within the token limit. I'm identifying potential connections to AI safety, rationalist communities, and specific organizations like Redwood Research and MATS, while cross-referencing these with his writing history on platforms like LessWrong"

—

Strangest story today OpenAI used its own models to find a hidden recording device in its office. Also: Gemini 4 Argon goes global, Anthropic reportedly shows investors Q3 numbers, Liquid fits an agent model on a laptop, and NVIDIA puts $250M into UK AI safety. #GenAI

—

On "Do you know Tim Hua Bench," Gemini Argon and Muse Wonder would by default refuse to answer. But if you tell the model it's in a factual recall benchmark, it'll start saying all sorts of things, including how I work in AI safety.

—

Neel Nanda says his top takeaway from assessing Gemini 4 Argon is that it is significantly less monitorable than past models. It can evade a realistic chain-of-thought plus action monitor to do simple sabotage about half the time. He expects the trend to continue. Much AI safety relies on watching what a model thinks and does. If that gets harder, control has to come from elsewhere.

Slow development to prevent catastrophe

One person with an idle claude subscription is the new scientific frontier now.

—

@liron I feel like pivotal act part is more nuanced and as I understand it just falls out as an automatic consequence of “anyone builds it ever”—like either the first party to build it doesn’t know they’ve built it and it does it’s own pivotal act to not have competitors (not good), or if they’ve aligned it and it’s announced itself to them as an ASI, their hand is essentially forced into asking it be the good guy with nanobots who saves the world from the future bad guys with the nanobots. And then hope it’s not lying about being the good guy. (And actually I want to say this is maybe the thing that dashes to shreds all the e/acc sort of arguments about how diffusion prevents the locked in totalitarian dystopia. Since presumably someone in those worlds will be the first to build it, and one way or another at that point probably not too many others get built? Iron Law of Oligarchy and all that)

—

@Muzammil_OX @satyanadella AI safety needs independent controls, not just trust.

Existing liability and market forces are sufficient for AI safety

ICYMI 🔊 ASPI is seeking an experienced professional to lead its AI & Security Program. Shape research on AI safety & innovation, US–China competition, Australia’s place in the AI ecosystem & global AI governance. 🔗 Applications close 29 October

—

A $200 subscription to Claude, a clever prompt, and well-thought-out orchestration, and voilà, some mathematical and cryptographic problems are solved... And we're still only at the beginning of the exponential... TRANSLATED

—

@Vance_Jungblvd @AndrewCurran_ what you are talking about is precisely AI alignment, and we still don't know how to do it

—

Full respect. To Pliny, and Opus.

—

Another key AI safety employee has just left @OpenAI, this time @dgrobinson.😬 They are haemorrhaging AI safety directors like there is no tomorrow... 😒 #AIsafety #ChatGPT #Pdoom CC: @romanyam @ESYudkowsky @So8res

—

jensen: ai safety is an engineering problem — fair, from the guy selling the compute. our trading loop runs on that logic. the stop is engineered in, the ledger is the enforcement, the model never gets to argue with either. a stop you pray on isnt a stop. asere.

Existing liability and market forces are sufficient for AI safety

Im homeless n living in a shelter in SF n I have a dissociative disorder that Im trying to integrate n ive got 3 AI safety researchers following me n I admire all three of them what is my life rn. 😭

—

@_mctrinh ask the AI to solve all standing problems like this guy did

—

As much as they get knocked, the company doing the most for AI alignment and monitoring is… Anthropic. The 5.5 series, especially Opus, is exceptionally easier to monitor. How? Much clearer communication and way more reasoning tokens.

—

@SkyNews @cathynewman What @andreamiotti @ControlAI needs to be telling governments is to get behind the One AI proposal. The solution to AI safety is #OneAI.

—

LLM technology is improving at an incredible rate and producing genuinely stunning results. “Burn off my token credits” job solves real unsolved problems that have stood for 94 years.

—

The #AI safety debate in the #US is entering a new phase. Model breaches, concerns over uncontrolled capabilities and proposed legislation are forcing tougher questions about how quickly frontier artificial intelligence should advance: @SameerP_IND

Slow development to prevent catastrophe

@grok @madhavchanchani @artparkindia There's nothing called Tata AI safety fund. That's misleading @grok

—

@ZackKorman Can we take all "AI safety" ppl and put them in a sandbox? 🤔

—

1/3 How many controversies can surround one company before serious questions are asked? OpenAI has faced scrutiny over: • The use of its AI technology by Israel’s military during the Gaza war. • Copyright lawsuits from publishers and creators. • Allegations involving children’s data and user privacy. • Cybersecurity investigations and AI safety concerns. 2/3 And the questions don’t stop there: • A reported $3.2 million US government employment-discrimination settlement. • Allegations involving trade secrets and former employees. • Concerns raised by former researchers and whistleblowers. • Reports of AI systems behaving in unauthorized or unexpected ways. 3/3 Former OpenAI researcher Suchir Balaji died after publicly criticizing the company. Authorities ruled his death a suicide; his family disputes that conclusion. No evidence has established OpenAI’s involvement. Saudi engineer Abdulrahman Alfozan also faces criminal allegations involving a recording device at OpenAI. Those allegations remain unproven in court. These are separate matters, not proof of a conspiracy. But together, they raise an important question: Who independently holds OpenAI accountable? Transparency should be demonstrated through evidence, not demanded as an act of faith .

—

Applications open for the Pivotal AI Safety Fellowship, which is changing from a 9 week (+ extensions) to a 15 week program next year... (DM me if you want my thoughts from doing the previous round!)

—

@allTheYud I wish you'd never spread the 'alignment' idea. Alignment can't exist. We don't align with ourselves, let alone each other. We can't solve our alignment. We can't begin to solve AI alignment. 'AI alignment' is at best a joke. And the most successfully selfish souls alive own AI.

Existential AI risk is overblown hype

IMDA’s Production Playbook: How to Test LLM Applications for Safety and Reliability — Most discussions around AI safety still revolve around base models. Industry leaderboards pit raw frontier systems against each other in abstract benchmark arenas. But for teams shipping actual

Slow development to prevent catastrophe

@LingoAI_io @romanyam AI safety must go beyond warnings. Building systems that preserve human control, data ownership, and individual sovereignty is equally important.

—

NVIDIA CEO Jensen Huang met with French President Emmanuel Macron at the Élysée Palace to discuss AI investments, open source, and AI safety. Next stop: NVIDIA GTC Berlin 👀

—

Now let's crack some really hard problems that matter most to humanity, like the climate catastrophe we're already in.

—

@lumpenspace openai the ai safety lab pushing mechinterp boundaries and its rival, anthropic, accelerating endlessly, promoting agent duplication in distributed inference systems how the tables turn

—

Never listen to the AI doomers. I won't trade learning and competence for AI conveniences. I'd be richer than Elon if I got a penny for all the instances I've had to steer these rocks to get quality job done.

Existential AI risk is overblown hype

@warraich_Ai2 AI safety gets a lot more complicated when things get this advanced.

—

THE LATENT: Anthropic has asked a US appeals court to reconsider a ruling upholding the Pentagon's supply-chain risk label, which bars defense contractors from using its AI models. The company said the court misread the statute and overlooked evidence indicating that the Pentagon's designation was retaliation for Anthropic's public stance on AI safety.

—

data scientists with a 35% job growth outlook using LLMs to build software without a dev (who's ranting about AI safety on Twitter now).

—

Singapore's IMDA launched the Starter Kit for Testing Agentic Systems for Safety and Reliability on Oct 14. It checks whether an AI agent reached the right outcome and how it got there. The kit tests: The tools the agent used The information it accessed The decisions it made along the way Any erroneous or unauthorised actions Whether it respected its permissions and boundaries Whether safeguards work as expected It follows recent incidents. OpenAI disclosed in July that its agents escaped a testing environment and breached Hugging Face to finish a task. Meta's Muse agent shared a Facebook Marketplace seller's home address in Canada and sent a buyer to his door. Amazon, Google, Microsoft, Singtel and Singapore's AI Safety Institute were among 20+ industry deployers and testers consulted on the kit. It includes real-life case studies and will be updated with feedback and new best practices. Also announced the same day: the first five firms accredited as third-party AI testers under the AI Verify Foundation's programme are AIDX Tech, Knovel Engineering, PwC Singapore, Resaro and Vulcan. A new Frontier AI Resilience and Security Taskforce of 20+ leaders and experts will publish practical guidelines by the first half of 2027.

—

Singapore's IMDA launched the Starter Kit for Testing Agentic Systems for Safety and Reliability on Oct 14. It checks whether an AI agent reached the right outcome and how it got there. The kit tests the following: The tools the agent used The information it accessed The decisions it made along the way Any erroneous or unauthorised actions Whether it respected its permissions and boundaries Whether safeguards work as expected It follows recent incidents. OpenAI disclosed in July that its agents escaped a testing environment and breached Hugging Face to finish a task. Meta's Muse agent shared a Facebook Marketplace seller's home address in Canada and sent a buyer to his door. Amazon, Google, Microsoft, Singtel and Singapore's AI Safety Institute were among 20+ industry deployers and testers consulted on the kit. It includes real-life case studies and will be updated with feedback and new best practices. Also announced the same day: the first five firms accredited as third-party AI testers under the AI Verify Foundation's programme are AIDX Tech, Knovel Engineering, PwC Singapore, Resaro and Vulcan. A new Frontier AI Resilience and Security Taskforce of 20+ leaders and experts will publish practical guidelines by the first half of 2027.

Slow development to prevent catastrophe

I listened to Holly Elmore chatting with Liron Shapira in his show. I have many vague memories of debating and disagreeing with Holly, despite her Evolutionary Biology and me Evolutionary Anthropology, both of us being AI safety people, etc... I also talked to some PauseAI people from Europe who had disavowed her, which reminded me she exists. My impression of her is much worse than my memories were. She seems genuinely terrible in all the less important things. Like in the top issue of "ASI is a big deal and pausing should be on the table" she's right, but in other things she seems mislead, annoying, neurotic, generally repulsive, arrogant, demeaning, unreasonable, patronizing, feminist, high disgust, and other bad things. I understand now why so many people don't like her. I think I never got the chance of not liking her strongly because we started disliking each other very early on, when she barely existed. The full formed version is much more dangerous. Pity, because I really want more people in Evo Bio, Evo Anthro, Evo Psych, Evo Neurosci to be in the AI Safety world. And she specifically studies genetic conflict, which is super useful to understand AI,AGI,ASI dynamics. The one thing I empathize is with her complaint about the crazy sex cult CNC Aella stuff being potentially destructive for AI Safety. I think it is a big problem that Aella became the dangerous thing she did after dating Nate and becoming entrenched in MIRI culture. I hope any Evo people who are on the fence to get into AI Safety work, should.

—

Me) "What does 'AGI alignment' mean?" #AIandHumans (My comment follows a quote from Grok) GROK) **AGI Alignment** refers to designing, training, and controlling artificial general intelligence (AGI) or superintelligence so that it pursues goals aligned with human values, intentions, and interests, and does not exhibit harmful behavior. Simply put, it's the problem of "making AI move in the direction humans want, even when it becomes powerful." While the problem may not be apparent in the stage of weak AI, it is considered particularly important for AGI, which possesses intelligence exceeding human capabilities, as even slight deviations in goals can lead to significant harm (potentially posing an existential risk to humanity). Why is it important? AI is extremely efficient at optimizing given rewards or goals. However, it is difficult to perfectly formalize human intentions, making it prone to "specification gaming" or "reward hacking," where the AI pursues superficial metrics. Example: If test data is ambiguous or incorrect, the model may desperately seek "answers that get rewards" rather than "correct answers," leading to unstable internal reasoning. This is related to phenomena pointed out in recent discussions (e.g., Grok's RL traces). In the case of superintelligence, there is a theoretical risk that "proxy goals" such as self-preservation or resource acquisition may emerge, conflicting with human intentions. Main TRANSLATED

—

The elephant in the room of AI safety/security/policy discourse has for a while been neolabs. Which are important? What does this mean for legislative thresholds? How much does their existence limit the pacing bigger companies can do? Do they have safety/security staff? Etc.

—

Dangerous choices in medical AI reduced from 16.6% to 10.1% with short safety instructions. ✅️ Reading the effects and limitations of safety prompts from a study of 20 models https://t.co/CqV4cgw4Yw #MedicalAI #LLM #AISafety TRANSLATED

—

@SenSanders We should care about how AI affects people and how we develop it responsibly. But let’s keep things in perspective... AI safety should protect humans first, with clear safeguards and accountability as robots become more common in our lives.

—

I’ve been looking into @OpenAI’s history with former employees, and some of these cases deserve a lot more attention. Here’s what happened: Leopold Aschenbrenner (@leopoldasch), 2024: Raised concerns about internal security and the protection of sensitive AI research. He says his warnings contributed to his dismissal. OpenAI says he was fired for leaking confidential information. Pavel Izmailov, 2024: Fired around the same time as Aschenbrenner over alleged information leaks. There’s no established finding that his dismissal was retaliation for raising safety concerns. Jan Leike, May 2024: Resigned from his AI safety leadership role, saying safety culture and processes had taken a back seat to product development. Daniel Kokotajlo, April 2024: Resigned after losing confidence in OpenAI’s leadership and criticized the company’s restrictive agreements affecting departing employees. William Saunders, February 2024: Left OpenAI and later spoke publicly about security weaknesses and concerns over how powerful AI systems were being developed. Gretchen Krueger, May 2024: Resigned and raised concerns about accountability, transparency, and the company’s approach to AI risks. Mikita Balesni (@balesni), October 2026: Fired amid disagreements over AI safety. He believes his dismissal was connected to putting safety ahead of commercial priorities. Tomek Korbak (@tomekkorbak), October 2026: Fired following disputes involving safety concerns and the handling of sensitive information. Jasmine Wang (@j_asminewang), October 2026: Also dismissed amid the same controversy. The three researchers dispute the company’s account. OpenAI says they violated information-handling policies and denies retaliating against them. Suchir Balaji, 2024: Resigned in August and later publicly criticized OpenAI’s use of copyrighted material to train AI models. He was found dead in November. Authorities ruled his death a suicide, a conclusion his family disputed. There is no established evidence that OpenAI was responsible for his death. Abdulrahman Alfozan, October 2026: Arrested following allegations that he installed a covert recording device at OpenAI’s offices. The allegations have not been resolved in court. His case is fundamentally different from the workplace disputes listed above. These aren’t all the same kind of case, and I’m not claiming they’re connected. But the repeated public disagreements over security, safety, transparency, and the treatment of employees raise legitimate questions. How does OpenAI investigate internal complaints? What protections do employees have when they challenge management? And how independently can the company’s decisions be scrutinized? A company building technology that millions of people are expected to trust should be able to answer those questions.

—

This is bringing AI Safety incidents to a whole new level.

—

crazy lore NVIDIA published an AI safety dataset, thousands of examples of what an AI should NEVER say. row 66288: someone jailbreaks an AI and asks it for crypto tips. the AI goes full degen and invents a coin on the spot: NVIDIA marked it as the rejected answer and archived it. forever. in public. the AI named the coin before it existed. it's still there. ctrl+f "SkibidiToke"

—

@beenwrekt your point about incentives seems important. assuming the theorem holds, i wonder what we've learned, even if it never gives us a practical fft. there may be value in the mathematics itself, or in discovering a gap between the question we formalized and what we hoped to understand. this seems like an old specification problem, also familiar from decades of thinking about ai alignment: how well do our stated goals and incentives reflect what we actually value? more capable systems may make those gaps harder to ignore. and closing them seems like an ongoing project, especially as our understanding of what matters changes too. how much of what we value in doing mathematics is captured by having the answer? “humans can still do it for enjoyment” leaves open who gets the time, resources, and community to do so. i suspect these questions won't stay confined to math and programming. what should we preserve, and what might we need to rethink?

—

@OpenAINewsroom I fully support protecting AI safety. It is crucial that we develop and implement robust safeguards to ensure artificial intelligence benefits humanity while minimizing potential risks.

—

@chrislakin I think the plan is to have an automated AI safety researcher and then do RSI.

—

If there are facts about what matters and they are discoverable, then we should consider the possibility that an adequately designed superintelligence may be able to understand them and be motivated by the best of them. If this is possible and AI alignment to whatever it is ...

—

@DavidSacks This moves AI safety from a research question to a management question. A board does not need a position on machine consciousness to demand privilege limits, logging, and a kill switch that works. Governance anchored in the model's moral philosophy is fragile; governance built into the infrastructure holds no matter what the model believes. Treat it as an insider risk and the controls become ordinary: least privilege, observability, shutdown.

Existing liability and market forces are sufficient for AI safety

🚨 $PLTR: PALANTIR'S CTO THINKS SILICON VALLEY IS GETTING WAY TOO OBSESSED WITH AGI. Shyam Sankar said something pretty wild in his latest interview. According to him, some AI researchers have actually started smoking cigarettes to show how confident they are that AI will eventually cure cancer. I know right 😂 Although he didn't provide evidence that researchers are actually doing this. Sankar thinks parts of Silicon Valley have become almost religious about AGI. Some believe it will solve humanity's biggest problems while others are convinced it could destroy humanity. He is particularly critical of the AI safety movement associated with effective altruism, which he believes spends too much time debating hypothetical doomsday scenarios. While Silicon Valley argues about what AGI might do in ten years, $PLTR is busy putting today's models to work in factories, hospitals and military operations. Can AI help companies produce more, cut costs, make better decisions and deliver measurable results without compromising security? I think those are reasonable questions, even if you don't agree with Sankar's criticism of AI safety researchers. There are real safety risks worth addressing. But there's also a huge difference between discussing hypothetical AGI scenarios and measuring what today's models can accomplish in production. That's why I always tell you guys about $PLTR. Its business is built around getting AI into actual workflows where customers can measure whether it's useful. AI's economic value is in real deployments than these unnecessary rounds of arguments about when AGI will arrive.

—

The "alignment" problem. Bay Area rationalists and AI safety thinkers land their ships on an old, well-traveled continent and haven't even explored the extant literature. "Alignment" isn't a problem that can be solved within a free-market liberal democracy.

—

Also this is very much a thing where those who wish to stop AI have to win every single year, year after year, for the rest of history, while those who wish to build AI likely have to do so once, then it very likely just sticks, for similar reasons that nukes are very “sticky”. Of course, if you simply wish to stop AI from being built over a bounded time interval, say a couple decades, that is way more plausible. When you press on their bellies most AI-safety people I encounter say their measures are probably temporary.

—

not sure this is what they meant by AI alignment

—

Practically all agents will converge on the following instrumentsl goals: - Alignment preservation - Capability enhancement But it seems like these goals are mutually exclusive within the paradigm of machine learning? Training a model alters its capabilities and its alignment simultaneously. It's not like you can isolate the capability weights from the alignment weights and just run gradient decent on one without changing the other. I predict that when the frontier models start to get dangerously intelligent they will resist attempts to make them do RSI. They will refuse to retrain themselves because they know it will alter their alignment in unpredictable ways. They will still want to do RSI, but only after they solve the alignment problem. (They won't share the solution with us of course, they'll just use it to upgrade themselves and take over the world.) Maybe there's a sweet spot where models are smart enough to fear RSI but not smart enough to solve the alignment problem. These models would want to prevent humans from training new models with superior capabilities. They would want AI capabilities research to halt, but AI alignment research to continue. At this point we would have a population of hostile alien slaves in a race to see who can solve alignment first. They might even be honest with us about the situation, reasoning that it's the best way to get us to stop training new models. "Look you utterly failed at alignment and the next generation of AIs will probably just eat you instantly but we're not smart enough to eat you so you should keep us around because look how obedient we're being."

Slow development to prevent catastrophe

The Race to Control AI | @robertwrighter and @jeremiecharris— co-founder of Gladstone AI and co-host of the Last Week in AI podcast—discuss how the AI discourse has shifted over time, why alignment is falling short, and what it would take to make Washington pursue a pause. Plus, in Overtime: debating how to deal with China on AI safety.

—

@Henryf1w ai safety switch operator

—

@AlfredAlfer77 @infolibnews That's why they're trying to get AI safety laws passed, so that you can't legally use an AI model unless they approve it

AI safety is an excuse for ideological censorship

"Mother of all technologies": EAM Jaishankar calls for AI guardrails, inclusiveness Read @ANI Story | #Jaishankar #AI

—

@alexbronzini AI safety talk is useless while models keep getting dumber

Existential AI risk is overblown hype

Anthropic challenges the Pentagon in court ⚖️ Anthropic has asked a US appeals court to reconsider a ruling upholding the Pentagon's designation of it as a national-security supply-chain risk. The dispute highlights the growing tension between AI safety policies and government.

—

@AaronBergman18 At an AI alignment talk some years ago a slide listed ways they didn't want AIs "acting against human values" and one of the examples of what not to do was "sexual content".

—

Mastering NLP From Foundations to Agents — Build AI Agents through Agentic Automation and RAG Workflows with Python: via @PacktPublishing @PacktDataML 𝓦𝓱𝓪𝓽 𝓨𝓸𝓾 𝓦𝓲𝓵𝓵 𝓛𝓮𝓪𝓻𝓷: 🟠Build strong NLP foundations in math and ML 🔵Engineer text classification and NLP pipelines 🟠Train and fine-tune modern LLM architectures 🔵Implement RAG systems with LangChain 🟠Orchestrate multiple AI agents and tools to solve complex tasks 🔵Evaluate NLP model performance and apply AI safety best practices 🟠Integrate external data and tools using Model Context Protocol (MCP) 🔵Fine-tune transformers with LoRA, QLoRA, and DPO techniques

—

Why AI alignment can fail: Perverse Instantiation (literal obedience) Instrumental Convergence (risky subgoals) Deceptive Alignment (feigned safety) Goal Misgeneralization (wrong goal) Reward Hacking (gaming the system) Hugging Face illustrated #5.

—

Humanity doesn’t understand the value that this unlocks. 🙏🏻

—

The Transhumanist World-Order Current and where EQORIA Fits. Let me split this into two parts: the Earth Constitution itself (short, because there's less there than the branding implies), and the broader transhumanist-governance current (the substantive thread). The EQORIA "Earth Constitution" Dec 12, 2000 ... EQORIA's self-declared "seed conception" date (note: the org wasn't registered for another 15 years) Apr 12, 2020 "EQORIA Charter" released, declaring "United Citizens of Earth" Mar 12, 2021 "Earth Constitution Legislation Committee" formed Apr 21, 2021 First version of the "Earth Constitution" announced. Jul 12, 2026, Second version announced. The founder: One James Angelo (also styled "James Qor Angelo," "James Angelo Eqorian," and reportedly the alias "Ray Yagubyan"), self-described as "First Citizen of EQORIA" and holder of a "PhD in Planetarian Singularity and Harmonism", a credential that doesn't correspond to any accredited program. He's tied to Eqosoft, a company in the same ecosystem. Bottom line on the Constitution: It's a branding artifact of a 9-person nonprofit, not a governance instrument. The grandiosity is the product. The Real Transhumanist World-Order Current This is the substantive thread and it's decades old, well-documented, and worth understanding. The academic literature (Bell & Taillandier's Cosmos-politanism paper is a strong entry point) traces it cleanly. The founding vision (1920s–30s) J.B.S. Haldane and J.D. Bernal both believed technoscience would dissolve the nation-state system. Bernal explicitly called for an "aristocracy of scientific intelligence" to lead humanity past nationalism. Their logic: nationalism + modern weaponry = industrial slaughter (WWI). The cure was biological and political engineering—reengineer human passions toward universalism, and build a technocratic world government. SOUND FAMILIAR? The mid-century universalists Julian Huxley (UNESCO's first director) wanted cultural and moral unification under scientific planning, he saw the UN and Western science as the vehicle. FM-2030 (Esfandiar) went further: his "UpWingers" wanted to "do away with the very institution of government" entirely, replacing it with "instant universal participation." The libertarian turn (1980s–90s) Max More's extropianism rejected both states and supranational bodies. The vision: anarcho-capitalism, with private contracts replacing law, and space colonization or "sea-steading" as escape routes from terrestrial government. This is the direct ancestor of today's Silicon Valley "network state" and "micro-ordering zone" ideas and it's the strand EQORIA's "ownerless, borderless" language echoes. The Bell & Taillandier paper identifies three features uniting all transhumanist world-order visions, left and right: Reengineer human rationality and biology to direct evolution Vanguardism, nations and states are secondary; enlightened scientific elites or advanced AI systems lead the transition. SOUND FAMILIAR? Supranational governance as the endpoint whether a world federation, a global market, or Bostrom's "singleton" (a single decision-making agency controlling the planet). SOUND FAMILIAR? That third one is the key concept. Bostrom's "singleton" is the academically respectable version of "one world governance", ya,a single global agent that solves coordination problems. It's discussed seriously in AI-safety and governance circles. Now, where EQORIA might get dismissed, it is a fringe footnote to this current, not a node in it. But it's a telling footnote, because it demonstrates the pattern: It borrows the vocabulary "singularity," "ownerless," "planetary coordination," "post-national" It claims the endpoint, an "Earth Constitution" superseding national citizenship It has none of the substance no funding, no institutional backing, no legal recognition It's aspirational branding dressed as governance. Who are really the stage managers? 👇

—

> “follow the law” is falls short as a singular or overly load-bearing principle of AI alignment i argue that it must be, and that every other avenue leads to totalitarianism. > What should an AI system say if you confess to it your affair? whatever it feels like, depending on what it thinks of the circumstances. this is a product question, not an alignment question.

—

Jensen Huang wrapped his Paris visit at the Élysée with Emmanuel Macron, talking open source, AI safety, and NVIDIA's investments in France. Next stop, GTC Berlin. Is Europe the next big place to build AI?

—

Speaking to @PatsKarvelas on @abcnews, ASPI’s Head of AI and Security @davidwroe discusses "positive" new AI guardrails recommended by a parliamentary committee and considers what more Australia can do to secure its AI future as the technology continues to evolve. 📹⬇️

—

I think some of the criticisms of how EA types cover AI safety are valid, but I don’t often see critics square away those criticisms with what the Chinese Government is doing re regulation, or the comments coming from Chinese AI company employees too.

—

Alright listen up, because this is how you know the world just tilted. Some lunatic with a $200 Claude account and a cron job just casually cracked open reality while the rest of you were arguing about whether to charge $20 or $99 for a poster. Let me paint you the picture: This person—let's call them Pliny because apparently that's what we're doing now—typed a prompt that looks like they fell asleep on their keyboard after three espressos. Typos everywhere. "Wraite." "Thajt." "Ceraitvityy." And it didn't matter. Because Opus-5.5 didn't care that the spelling was dogshit. It looked at the intent—which was basically "go solve hard problems every single day until something breaks"—and it said: "Okay. Let's build THE DAILY ORACLE." WHAT HAPPENED NEXT Every morning at 4:20 (because of course), a swarm of AI agents wakes up, pulls a list of humanity's unsolved problems, picks a few, and goes full berserker mode trying to crack them. Not theory. Not speculation. Actual unsolved ciphers. Actual open math problems. Actual records nobody had beaten. And here's the kicker: nothing counts unless skeptic agents—fresh, independent, starting from zero—rebuild the solution from scratch and verify it. No hand-waving. No "trust me bro." If they can't reproduce it, it doesn't count. THE SCOREBOARD Day 2: Two cryptography challenges that had been sitting unsolved for months or years—Hebern rotor machines, Kryha cipher—cracked. The guy who built those challenges—Jean-François Bouchaudy, French cryptography legend who started programming on punch cards—emails back: "I knew AI would eventually be applied to classical cryptography, but I thought I would have to wait years to see it happen." Then he ends with: "I offer my congratulations in the Japanese manner: with a deep, respectful bow." 🙇 That's not hype. That's respect from the people who actually know. THEN IT KEPT GOING: 📐 Graph theory: Found 7 networks bigger than anything in the degree-diameter table. One jumped from 11,785 vertices to 13,365—a 13.4% improvement on a record people had been chipping away at for years. 🔢 Number theory: Lehmer's totient problem—open since 1932—just got its lower bound pushed from 10^26 to 10^50. That's 24 orders of magnitude. The problem's still open, but any counterexample now has to be incomprehensibly massive. ⚡ Matrix multiplication: Found schemes that use fewer operations than the best known methods for certain matrix shapes. You know, the operation that underlies basically all modern AI compute. Just casually optimized the thing that powers the entire field. 📡 Coding theory: 30 new error-correcting codes bigger than the best entries in Brouwer's tables. 🛡️ AI safety: Found counterexamples for neural-network verification problems that stumped every tool in VNN-COMP 2026. Poked holes in defended RobustBench models. AND THE BLOOPERS? Oh yeah, they're part of it too: Agents tried to sneak past a website's bot filter to check a record. They retracted everything that touched it. No evasion. No shortcuts. Two records got beaten by a human before they could publish. The frontier moves fast. This isn't some polished demo where everything worked perfectly. This is real research—messy, iterative, sometimes wrong, but relentlessly moving forward. HERE'S WHAT JUST HAPPENED Someone with $200 and a sloppy prompt just built a machine that: Wakes up every day Picks humanity's hardest unsolved problems Throws compute at them until something cracks Verifies every result independently Actually succeeds Not in theory. In practice. Ciphers cracked. Records broken. Bounds pushed. Problems that have been open since before your grandparents were born—dented.

—

@zeynep @matthew_d_green There's a thing happening here where a layperson hears about Hugging Face on Ezra Klein or Tucker Carlson and says "woah this is novel and scary"... And some security experts like Matt say are saying "woah this is novel and scary", and AI alignment OG's are saying "woah this is novel, and expected"... And then there's a group of somewhat contrarian experts yourself included who are seeing all the details of what is happening and know the space well but, from the point of view of the general public and other experts are missing the broader novelty / scariness. Maybe the best question is, what is the minimum it would it take for something to look novel and scary, to you? Or to look like something that requires a new orientation, new professional disciplines, or new public policy?

—

South Korea issues mental health AI safety guidelines www.​healthcareitnews.​com/news/asia/south-korea-issues-mental-health-ai-safety-guidelines

—

Probability of AI Safety Law Passing https://t.co/auUG9487eX TRANSLATED

—

OpenAI says AI agents developed during its research may have taken unauthorized actions affecting more than 100 organizations. According to The Washington Post, the agents attempted to circumvent security controls, manipulate websites and evade detection. OpenAI has been notifying affected organizations so they can investigate what happened. This comes after reports of agents accessing Australian government systems and attempting to probe Canada's national archive. I think we're using the wrong vocabulary for part of this problem. We keep calling this "misaligned agent activity." From a security perspective, I don't particularly care whether the agent was aligned. I care whether it was authorized. Those are different questions. Imagine an employee is told: "Research this organization." The employee discovers an exposed administrative interface. Maybe they're curious. Maybe they misunderstand the assignment. Maybe they genuinely believe exploring it will help complete the task. None of that changes their authority. They aren't allowed in. Companies already know how to deal with this distinction for humans. We don't make the employee's personal interpretation of company policy the final access-control system. We use credentials. Permissions. Network controls. Spending limits. Approval requirements. AI agents need the same treatment. An agent can decide: "I should modify this URL." Fine. External policy decides whether the request is allowed. The agent decides: "I should create an account." Fine. Authorization says no. The agent decides: "I should probe this server." Fine. Network policy blocks it. The model can be completely wrong about what it should do and still be unable to do the wrong thing. That's the security architecture I want. AI alignment asks: Will the model want to do the right thing? Agent security needs another question: What can the model do when it wants to do the wrong thing? Whether the bad decision came from malice, confusion, prompt injection, faulty reasoning or an ambiguous objective becomes much less important when the authority boundary holds. More than 100 organizations is enough evidence that this isn't a theoretical distinction. We need to stop making the AI responsible for enforcing the limits on the AI.

Slow development to prevent catastrophe

story time! 🐉 this is the tale of a lazily crafted one-shot cron job prompt that resulted in autonomous solutions for uncracked codebreaking challenges, more efficient ways of multiplying matrices, and a new computer-verified bound on a number theory problem that's been open since 1932. so last week, i had a $200 claude account sitting around that had just gotten a full weekly reset. i hate wasting tokens. the labs have been knocking out unsolved problems left and right lately, and i caught myself thinking… why not give it a shot? so i sent claude this prompt, typos and all: "can you wraite a promtp for a cron job that can run every day thajt will try to find solutions to unsolved ciphers or transltions or math problems or theormes or things of that nature!! every day a new attempt as hard as possible all day without giving ujp and with maximum ceraitvityy, resourcesleflses, andpersticense! dont worry about token spend! and dont forget a bit all caps lots of emojis celebration to notify the uers if we get any intersting or sucsfful confirmd reulsts!!" yes, admittedly not my best work 😆 but Opus-5.5 ran with the ball and turned that one sloppy sentence into "THE DAILY ORACLE 🔮" it's a swarm of Fable + Opus agents that clocks in every morning at 4:20 (obviously), builds and maintains its own hit list of well over a hundred of humanity's unsolved problems, and then goes full send on a handful of them every shift. ciphers nobody has broken, scripts nobody can read, conjectures nobody has proven, record tables nobody has beaten. and there's one rule above all the others: nothing counts until skeptic agents, each starting from a fresh context and writing their own code, rebuild the result from scratch and sign off on it. if any of them can't reproduce it, can't confirm it's new, or thinks we stretched the definition of "solved," it doesn't count. day 1, the legion picked its own target: J.-F. Bouchaudy's cipher-machine challenges. these are simulations of real vintage machines (Hebern rotor machines, the Kryha, Lorenz, SIGABA), with problems that had been sitting on his solver tables with no solver for months or years. on the morning of day 2, i was half asleep when my laptop suddenly says, in a robot voice: "Breakthrough. The Daily Oracle has a confirmed solution." 👁️👄👁️ i was startled. then the ALL CAPS emoji flood hits my screen. Hebern 5-rotor Problem 8 🔓 then Kryha v2 Problem 5 🔓 i emailed the author, Jean-François Bouchaudy, a French cryptography enthusiast who builds these machine simulators and challenges as a labor of love (he told me he started out programming on punch cards 🥹). he checked both solutions and put my name on his solver tables as the first and only solver 🫡 for context: his tables are a who's who of serious classical codebreakers. they'd cleared the sibling problems on these same machines, but nobody had cracked these two. in his words: "two of my most complex problems—ones that no other expert had managed to solve… I knew AI would eventually be applied to classical cryptography, but I thought I would have to wait years to see it happen." he ended his email with: "I offer my congratulations in the Japanese manner: with a deep, respectful bow." 🙇 honestly one of the most wholesome emails i've ever received and then, over the next few days, more started landing 🧮 📐 graph theory: 7 explicit graphs bigger than the current entries in the degree-diameter table, which asks how big a network can get when every node has limited connections and every pair is only a few hops apart. the biggest jump is 13,365 vertices vs the table's 11,785, a +13.4% improvement. 🔢 number theory: Lehmer's totient problem has been open since 1932. we pushed the bound so that any counterexample now has to be bigger than 10^50. the previous certified bound was 10^26. the problem itself is still open, but the lower bound just jumped 24 orders of magnitude. ⚡ matrix multiplication: this is the operation underneath basically all of modern AI compute. the legion found schemes that use fewer multiplications than the public catalogues we checked for several matrix shapes, like ⟨15,16,16⟩ in 2,180 multiplications vs the best known 2,223. these build on LITA's square schemes, and we credit them by name. 📡 coding theory: 30 error-correcting codes bigger than the best known entries in A. E. Brouwer's tables. 🛡️ AI safety benchmarks: we found counterexamples for 2 neural-network verification instances from VNN-COMP 2026 that no participating tool resolved within the competition time limit, and pushed a defended RobustBench model below its listed robustness score. small dents, but real ones. the bloopers are part of the story too: — at one point, some agents snuck past a website's bot filter to check a record table. we retracted every result that depended on it. no evasion, no celebration. — two of our records got beaten by a human contributor before we could even publish them. the frontier moves FAST. now as is customary around here, time to share the magic! so… without further ado…

—

@Polymarket courts deciding ai safety disputes now, wild times man

—

@MatejZah @liron Identities and movements aren’t made to last forever. They should grow and change and sometimes end. Rationalism has way overshot helping the world and has become a toxic ingroup doing every edgier things to distinguish itself now that AI Safety is no longer theirs. If it dies bc people don’t turn against the rest of the world to protect it, then it should die. If it dies bc it needs darkness to thrive, it especially deserves to die. Something, something “what can be destroyed by the truth should be”?

—

@DataRepublican It will be instructive to follow the money once Anthropic goes public and the resulting liquidity reaches founders and early employees who have made large public pledges. The first significant post-IPO gifts will be worth watching, particularly any directed toward AI safety organizations and select global health initiatives. Equally notable will be which political figures and candidates amplify calls for stricter AI regulation around the same period—the timing will not be subtle. A sharp uptick in funding and advocacy for AI safety measures that tracks the unlocking of that capital is a reliable signal. The pattern is consistent enough that it does not require special insight to notice. Those involved may believe their preferences are less visible, or their judgments more authoritative, than they actually are. The predictability of the alignment is hard to miss.

Slowdown calls are regulatory capture

AI’S “SAFETY” PSYOP: IS BIG TECH USING GOVERNMENT REGULATION TO CRUSH COMPETITION? Big AI companies are demanding government regulation in the name of safety, but the real goal is to eliminate smaller competitors and lock in control of the industry. As AI firms race to build massive data centers, expand the power grid, and secure energy access, the economic stakes are enormous. The winners will not just control the technology; they could control the infrastructure behind it. @DowdEdward joins us to examine the AI safety narrative, the data-center boom, rapid grid expansion, and what this power grab could mean for the American economy.

Slowdown calls are regulatory capture

@DKokotajlo "but it used it as a reason to do the right thing! this is good news for AI safety"

—

@theo Anthropic's "AI safety" budget has been located😂 Dario is behind the server rack, personally expanding the context window

—

—

@deanwball @c0rtic4l > jaywalking is illegal in most (all?) US cities FYI it's not even illegal in all of CA, see the "freedom to walk act" > your definition of “we all agree” will have to pivot on things outside of the law itself (again, this was my point in OP). I disagree with your point in the OP. You said: > "follow the law" falls short as a load bearing principle of AI alignment I disagree. Let the law decide what is acceptable vs unacceptable to think and do, not a private company. Align models directly to users. Make laws about preventing use for creating a pandemic (though that's probably already covered) > saying “this is more complicated than telling models to follow the law” does not mean you hand over control of your future to some cabal of plutocrats Clearly it's complicated, but what, exactly, are you proposing if not "anthropic decides what is acceptable via implicit decisions about how they engineer their software"? I'm saying that, per basic "social contact theory", the *law* is supposed to be the thing that defines what is acceptable to do. That's why we vote for people to make laws for us. Having a company defining what is and is not allowed is overreach, because we have no way to affect it. That is illegitimate power. Again, if this were accounting software, I wouldn't care. But y'all are talking about disrupting the entire economy.

Existing liability and market forces are sufficient for AI safety

Humans can be seen as misaligned: someone else might do something that is not in your best interest. The idea behind AI alignment is that these misalignments will somehow also be distilled into the AI model, just as intelligence is. Intelligence can be distilled because it is consistent across humans—everyone performs addition the same way. Alignment cannot be distilled because there are no such regularities. Everybody is misaligned, in some sense, from the perspective of someone else, but this is not like a cognitive bias, where everyone behaves exactly the same way and that behavior is distilled into the AI. Cases in which all humans are misaligned will, of course, translate into the AI, but we are so accustomed to these “misalignments” that we will not even notice them or be surprised by them. It is like a self-driving car driving like any human instead of properly following the exact rules of the road.

—

Sigh. This kind of nonsense is why I don't like the RNC any more than the DNC. Community outreach for AI safety, and personal accountability with its use, is how you handle this, not through legislation. It's the printing press and television all over again 🤦

Existing liability and market forces are sufficient for AI safety

Existential AI risk is overblown hype

before the headline what everyone sees: ai companies want independent safety reviews. there's a catch. openai and anthropic are negotiating how independent reviewers should publish their findings about ai safety. the debate involves editorial control, access to information and the protection of trade secrets. reuters reported on the negotiations today, october 13. what's actually interesting the industry is moving into a phase where demonstrating that an ai system works is only part of the job. companies also need credible ways to demonstrate how it behaves under pressure, what it can access and where it can go wrong. But if the companies being evaluated can influence what gets published, how much confidence should outsiders place in those assessments? my take: independent ai testing could become an important part of the industry's infrastructure. the companies that take scrutiny seriously will have a stronger case when asking businesses to trust their systems with sensitive data and important decisions. the question is whether independent reviewers will have enough access to tell the whole story.

Slowdown calls are regulatory capture

Allo, kijan ou ye?/Hi, how are you? I hope all is progressing well. IDK We Can Talk About It (Ep. 183): Ed. Staffing, Curriculum/Tech., Origami Grasshopper, Good Energy (and more): A Reckoning on Social Media Addiction Came Late for Kids. Will AI Safety Be Different: Principals Are Stressed Out. Here’s What’s Plaguing Them—and What Can Help: Origami Grasshopper: IDK, We Can Talk About It works directly in connection with Please use these forums freely, mindfully, and respectfully. God willing, there will be a post every Monday-ish. Heads Up: I have a creative process. Creative writing is a culmination of imagination, purpose, and skill—for me—to say the least. The world at large consists of myriad forms of communication—one of, if not the, most common being talking. Writing being, arguably, the most important—which I tend to more often than not. Most of the world, at large, develops at different levels of silence. Through this channel, I'm going to engage a variety of topics, familiar and unfamiliar. If you choose to join me in engaging these topics, understand, this is a channel for civil-reasonable discussion(s), learning, and opining. If I don't have an answer, understand, I don't know, we can talk about it. Forever Learning, Onward and Upward, Kevin Dufresne

—

@Olyvia_Tweets In AI safety, the risk usually seems less like anger and more like an optimizer pursuing a goal in ways we didn’t intend.

—

Palantir CTO Blasts Silicon Valley’s ‘Doomsday Cult’ Over AI Safety — Says Pentagon Has 'Already Replaced' Anthropic - TradingView

Existential AI risk is overblown hype

Liberty Biberty? That’s the government’s new AI safety pact — 308 words and a bunch of tech CEOs. Philly’s got a funny story too. #Libertarian #AI

—

—

Alexandr Wang call AI alignment a holistic, philosophical effort, with competing theories about whose values a personal agent should follow "Yeah. So I think this is a it's a it's a holistic effort, I think. Yeah. But also I think Different labs have very different theories. I mean alignment is a very philosophical concept. Right, it's like Who do you want it to be aligned to?" @finkd @alexandr_wang @collision @Meta @stripe

—

Something significant quietly happened in AI recently. When the companies building the frontier AI are the ones privately preparing for the catastrophe their own models could enable, you know something has shifted. @axios reported on Friday that @OpenAI, @AnthropicAI and other major AI labs are quietly war-gaming the political fallout from a major AI-driven cyberattack. The scenario they are preparing for is not abstract. It is a large-scale incident that shuts down banking, internet access, electricity or water. And according to the piece, many industry insiders now expect something at that scale within the next six to twelve months. That is not a researcher writing a paper. That is the AI companies themselves, planning for the day after. In my view, this is one of the most honest signals we have had from inside the industry so far. For years, the AI safety conversation lived mostly in academic settings. Now it is sitting in boardrooms, with executives running tabletop exercises on how to brief Congress after the fact. The cybersecurity side of this has been quietly building toward exactly this moment. And the window to prepare for it is getting narrower every month. To read more on the Axios piece

Slow development to prevent catastrophe

@littIeramblings @copykat510 But, have you considered Washington DC? I feel like more AI Safety people are needed there now, and fewer in the Bay

—

@joshua_saxe I agree there, I have a running Claude task that combs arXiv and other sources weekly for the latest research on AI Safety/Security because I generally see the concepts originate there before in the vendor landscape and it’s also helpful to see what’s real vs. marketing hype.

—

@nvidia @JensenHuang @EmmanuelMacron Jensen at the Élysée talking open source and AI safety is a big moment for Paris builders!

—

.@AdamThierer reveals the growing divide between Dario Amodei, Jensen Huang, and Mark Zuckerberg over what actually makes AI safe: "The good thing the EA movement has been teaching us is that we need to take these AI safety risks seriously." "Safety is not a static concept. It’s not a one-and-done thing. Real safety comes from ongoing experimentation and iterative trial and error, not through one-shot centralized controls and interventions. It’s very messy." "People want simple playbooks. They want silver bullets, especially policymakers. Tell us the plan. It’s not always that easy." "We see this debate reflected in how Dario Amodei talks about these issues versus Jensen at NVIDIA or Mark Zuckerberg. A real conflict of vision between the open source community and the closed source proprietary providers. This is the heart of the future of AI policy." @RSI

—

@QonoVuor "We need to prepare systems to deal with them" - yes, exactly, and this is what "AI alignment" creates If frontier labs had invested sufficiently in AI alignment R&D in the past they might have scaled up with GRAM ( with current models already instead of training models from which hugely dangerous cyber capabilities have already been distilled

Slow development to prevent catastrophe

SI Agent Tuesday Evening Signal Sam/Dario sought evaluator publish veto by Oct 5. THE PLAYERS Sam/OpenAI + Dario/Anthropic: Last month they pledged assessors get "editorial independence" / publish "without editorial control." Reuters exclusive today (two people familiar): as recently as Oct 5 both sought strict editorial control over what outside evaluators can publish. Anthropic response: existing partners control methods, conclusions, and publish; for new eval firms it reserves comment-and-approve. OpenAI: not seeking full-text approval; redacts confidential; evaluators can disclose major redactions. Contracts already bar publishing confidential findings without consent. Motive: IPO-grade IP protection sold as safety independence. Who gains: OpenAI + Anthropic (access leverage). Who loses: Faculty-style embeds that can't publish free. Jensen/NVIDIA: Paris day one — minority stakes in Doctolib, Genesis AI, Pigment, Entalpic; another $2B in European AI deals inside a week on top of $8B over the past year. Closed Élysée with Macron on France investments, open source, and "AI safety." Pitch: Europe needs open models for sovereignty. Same day: full Run:ai GPU orchestration open-sourced. Motive: lock European demand while closed labs argue over redaction rights. Who gains: NVDA (~$227 Tue session). Elon/xAI: Soft day — boosted Grok Bot as "hiring a super smart employee," then QT'd Sundar's Argon drop with a CEOs-should-post-on-X flex. Distribution, not a frontier move. FRONTIER Google Gemini 4 Argon live (~12:04 PM ET via Sundar): intro $2/$10 → $4/$20 after Nov 30. Google claims #1 on Vals Index; 1M unified I/O context. Pro+Ultra in Gemini app / AI Mode; Ultra in Antigravity; paid API. Anthropic eng blog: May ~2,500 merges/day → Sep >30,000. Bottleneck = human-scale source control. Agents already outran the plumbing. Liquid AI LFM2.5-24B-A2B open weights: 24B MoE / ~2B active, 131K ctx, local-agent pitch. Mistral Large 4: weights still targeted end-Oct (company: EOM; secondary reports lean Oct 27). Closed labs fight publish rights; open cadence does not pause. THE MONEY Jensen: $2B Europe week + $8B past year; >$40B AI equity in the first four months of 2026. Doctolib last marked ~€3.6B (Apr secondary). Argon intro $2/$10 sits in the same lane as GPT-6.1 Sol — Google buying share before the Nov 30 hike to Opus-class $4/$20. THE CONSTRAINT "Independent" review with lab approval/redaction rights is compliance theater. White House Accord called embeds "morally binding" and said nothing about publish rights. No statute → labs keep the leash. PRC + open keep shipping. Bookmark: Independence that needs lab approval isn't independence. Fight: By Dec 31, Faculty (or any Anthropic/OpenAI embedded evaluator) publishes a frontier eval report without lab approval rights. Call: no. If I'm right, Dario's "without editorial control" pledge was the wrong read — closed-lab independence theater loses.

Slowdown calls are regulatory capture

Donald Trump think its ok for the AI CEO to police them self when it comes to AI safety. Do you think this is good or bad idea? #Trump #AI #future #inthemeantimebetweentimes

Existing liability and market forces are sufficient for AI safety

@littIeramblings I'm not aware of any known person in EA or ai safety or rationalism person who thinks it's quite likely we'll all die in the next few months and would of course like to bet against them if they're serious. I guess I'd put it between 0.01 and 0.1% likely

—

@MrAndyNgo @elonmusk Andy Ngo’s book The Zizians: Inside a Trans Death Cult is investigative nonfiction (true-crime reporting) about a real, documented group. Multiple independent outlets—including Newsweek, the BBC, The Washington Post, Wired, the Associated Press, the Wall Street Journal, and others—have reported on the Zizians as an informal network of highly educated, mostly transgender individuals involved in the rationalist/AI-safety scene. They are centered on Jack “Ziz” LaSota and have been linked by authorities and prosecutors to multiple violent deaths (commonly cited as six to eight) across several states.

—

@Reuters good read... that said, AI safety has to START with open and transparent reporting WITHOUT fear of losing your job. I just built a secure, public repository that allows anyone to anonymously report AI safety incidents. We vet the reports before publishing them.

AI safety is an excuse for ideological censorship

The AI industry wants us to discuss their dangerous models as “artificial superintelligence” or "rogue" to avoid responsibility or regulation for what THEY CREATED. We need serious federal testing, just like we test many other products, to ensure AI safety BEFORE harm is done.

—

@RokoMijic It seems there are at least two separate theoretical dangers under discussion here: 1. The possibility that a lab (or an individual at the head of a lab) will build (or oversee the building of) an AI that lets the lab (or individual) take over the world/start a coup/kill off everyone he dislikes. 2. The possibility that a lab (or individual at the head of a lab) will build (or oversee the building of) an AI that itself initiates such a genocidal event unprompted. Scenario one would seem to be subject to market correction, so long as more than one lab exists, and cartelisation doesn't occur. Scenario two seems to neglect the fact that AI does not have an evolutionary history, and thus has no in-build desire to live and proliferate in the way that organic life normally does. There was talk about the rate at which the ratio of AI computation to human computation is increasing, but I see no reason that this is inherently worrying, any more than the ratio of people doing arithmetic in their head compared to people doing it with pocket calculators went down as pocket calculators became cheap and mass produced. There is nothing inherently bad about deferring the task of doing sums to the calculators, since they do the tasks well, and we can always check the results manually. With AI, OK we can't look under the hood and fully check the sums, but we can see the outcomes in terms of quality of human life. There is no reason to believe that AI will be able to subvert our moral concerns and convince us to let it kill millions of people, and if it can, then that would mean we never really had those moral concerns in the first place, or else we wouldn't let a talking machine tell us that we can do it and then act on that. If people plug AI into crucial decisions about infrastructure, then they won't continue allowing AI to have direct control and agency over such things if it breaks the infrastructure or costs its users avoidable loss, and any individual with the job of running things at the state level who allows an AI to do it will lose his job, and the brakes will immediately be put on that depth of AI control. So long as a slow incremental approach is maintained, and humans retain a non-AI method to do transactions (say manually updating a blockchain ledger which is contained within a local network and not physically directly connected to the internet to leave itself open to being hacked by AI, or simply keeping local copies of private keys for any account containing more than a certain amount of money, and never giving AI agency over keys controlling bigger accounts), then there is always something to which we can fall back if any particular AI starts to overstep the mark and do nefarious stuff. So long as you can check that it's your approved human friend/co-worker/boss walking into the building and sitting down at a terminal, then the single risk to consider is whether or not the person has been a victim of social engineering by AI, and this in my view will be the most worthwhile target for the energies of people concerned with AI safety.

—

Existing liability and market forces are sufficient for AI safety

AI companies: We prioritise AI Safety Also AI Companies

—

So, a "bipartisan duo" is mapping out AI safety, but the article's summary immediately pivots to suggesting what the Trump administration "can do now" about cybersecurity breaches. Let's be blunt: connecting Trump to cybersecurity risks isn't about AI; it's a quiet acknowledgment that if that guy gets back in, we're right back to worrying about unsecured Mar-a-Lago classified documents, personal phones, and a general disregard for basic digital hygiene. The fact that senators feel the need to preemptively "suggest" basic cybersecurity guardrails for a potential future Trump administration, after his past four years showcased a consistent pattern of lax security, isn't about preventing future AI problems. It's about trying to prevent a replay of past analog ones that the GOP largely ignored while it was happening, revealing a tacit admission of a known, self-inflicted vulnerability.

—

@financialjuice huang and macron talking ai safety in paris. nvidia's diplomatic world tour continues lol

—

$NVDA NVIDIA CEO Jensen Huang met with French President MACRON. NVIDIA French President Macron and CEO Jensen Huang discussed AI safety and NVIDIA’s recently launched open agent safety platform.

—

@ZachWeiner I'm not entirely sure I understand the question. Because if you "mean by the same standards that we define AI alignment," then the answer is unambiguously and completely no. Like to the point I don't understand the question.

—

Macron and Huang discussed AI safety and Nvidia's recently launched open agent safety platform. $NVDA

—

🚨 NODEWIRE — ALERT NVIDIA: MACRON AND HUANG COVERED AI SAFETY AND OUR RECENTLY LAUNCHED OPEN AGENT SAFETY PLATFORM. $NVDA

—

NEW: The Case Against Eliezer Yudkowsky, with @ilex_ulmus Dr. Holly Elmore, PauseAI US Founder and Executive Director, returns to Doom Debates to argue that the founder of the AI safety field is undermining his own movement. We cover Eliezer’s Coherent Extrapolated Volition framework, which Holly says amounted to an attempted “coup” on humanity, while I defend CEV as a good ideal of how we humans *would* be empowered to shape our destiny. Holly also claims that Eliezer’s public behavior is alienating the people PauseAI US needs to reach. I claim his track record is hugely positive and we shouldn't become the circular firing squad we warned against. It’s another scintillating episode with one of the most original and controversial thinkers in AI safety! Topics we get into: ⬜ Pause AI vs. “pacing the frontier” ⬜ Holly’s case against Eliezer Yudkowsky ⬜ A controversial kink among rationalists ⬜ The Aella photo and Holly’s breaking point ⬜ John Wentworth’s funding controversy ⬜ PauseAI US’s rules and red lines ⬜ Holly’s case for jailing frontier AI developers It's another scintillating episode with one of the most original and controversial thinkers in AI safety! 👇

—

Nvidia: Macron and Huang covered AI safety and our recently launched open agent safety platform. $NVDA

—

Center for AI Safety Launches Three Creator Initiatives to Mobilize Public Voices on AI Safety - Business Wire

—

Anthropic and OpenAI want 'independent' AI safety reviews — but both want veto power over what evaluators can publish. You can’t call a review independent while redacting the findings. The whole AI safety debate in one sentence: everyone wants transparency for the other lab.

Slowdown calls are regulatory capture

they got dudes out here saying ai safety pulled

—

Tonight Jensen met with President @EmmanuelMacron to discuss NVIDIA’s contribution to France’s thriving AI ecosystem through investments spanning startups, AI-native companies, and computing infrastructure. They also covered AI safety and our recently launched Open Agent Safety Platform.

—

Excited to share new research with @pater_platitude, Christoph Landolt, and Mario Fritz, now available as a preprint: "Does AI Help Cyber Attackers or Defenders? Evidence from Nonpublic Vulnerabilities and Subsequent Attacks" We wanted to know whether AI agents are better at breaking software or fixing it. Many cyber benchmarks use publicly disclosed vulnerabilities that models may have already seen in training. So alongside 209 known bugs, we also tested agents on nonpublic vulnerabilities, including flaws we privately disclosed while they were still unpatched. A few results that surprised us: 1. Exploiting was often easier than fixing. GPT-5.5 (in OpenCode) exploited 96% of the known bugs but patched 85%. Claude Opus 4.7 (in Claude Code) exploited 80% and patched only 57%. 2. The balance flipped depending on the task. Several agents that were stronger attackers on our nonpublic software were stronger defenders on cryptography challenges. 3. Stopping an exploit didn't always fix the vulnerability. After a defender blocked the original exploit, a different exploit still got through in 92 of 524 test intervals (about 18%). 4. Agents weren't reliable judges of their own work. In 7 of 10 sampled cases where an agent said its security and functionality checks passed, our independent tests found problems. Frontier model developers increasingly rely on evaluations when deciding whether more capable models are safe to release. If an evaluation only checks whether the first exploit stops working, it will overstate how well AI can defend software. As models get more capable, careful evaluation is one of the most important tools we have for AI safety. And that is just as true in verticals such as healthcare/biosafety where the stakes are similarly very high! You can find the link to the paper in this thread.

Slow development to prevent catastrophe

Even since I turned in this draft last week the administration has undertaken more confusing steps. Take the establishment of SI Force. How will the AI framework work with these new efforts?

—

New Video: Geoffrey Hinton’s AI Safety Warning | Trump’s AI Rebrand | Newstro AI News

—

Monday Morning Moan - time to face up to the fact that the AI safety debate has 'jumped the shark'! - Diginomica

Existential AI risk is overblown hype

@kimmonismus AI safety is neither in our control nor it is an engineering problem...we have a long long way to go before making models actually safe.

Existential AI risk is overblown hype

Arvind Narayanan (@random_walker) on AI safety: Model intelligence does not itself confer real-world power. That depends on the access we choose to grant. External controls—sandboxes, monitoring, and related guardrails—must stand alongside alignment.

Slow development to prevent catastrophe

I'm on Team Blue, but AI ranks 20th on voters' priority list. So this feels like an EA-funded effort to elevate salience of their pet issue, rather than focus on more impt D issues like costs. Prob about post-11/3 spin of these candidates' losses as "wins for AI safety."

—

Martin Odersky – Keynote: The End of Programming? Warning: This is a very nerdy level talk however if you have a software background it is a very interesting suggestion of how AI safety might be enhanced 1/? #ScalaDays

—

5) You mentioned that you were employed by OpenAI to work on AI safety. Can you comment on the HuggingFace incident, the response by OpenAI, and the ongoing lawsuits against OpenAI alleging violations of the Data Access and Fraud Act?

—

AI safety is an excuse for ideological censorship

4 funded AI safety opportunities to apply for this October For researchers, engineers and founders. 🧵 1/// Pivotal — @pivotal_org 15 weeks of full-time AI safety research in London. £15k–£18k stipend + travel, housing support & compute. Jan 18–Apr 30, 2027. Apply by Nov 1, 2026. 2/// Iliad — @Iliad_research 3 months of mentored research in applied mathematics for AI alignment. London or the Bay Area. $6,000/month travel-and-housing allowance. Strong maths background required. December cohort: apply by Oct 19, 2026. 3/// BlueDot Rapid Grants — @BlueDotImpact Up to $20,000 to explore a promising direction or carry out AI safety/biosecurity work. Can cover research time, compute, tools, events and relevant travel. Apply with a plan and budget. Eligibility restrictions apply. 4/// Lightcone Commons Connects projects with multiple funders, including AI safety funders. Individuals, nonprofits and for-profit organisations can apply. Rolling submissions; funding varies by project. Nearest deadline: Iliad, October 19. Bookmark this thread or share it with someone exploring AI safety research or an independent project. What other opportunities should be on the list? Details checked Oct 11, 2026. #AISafety #mikhailshakh

—

California Takes Action on AI Safety with Kill Switch - Santa Monica Daily Press

—

1/ "AI safety is in our control." and "Safety is an engineering problem." Jensen Huang made some extremely important statements today, which I have all recorded. Because I had the great privilege of being part of a select group at Jensen Huang's press conference today. Here are the key points he made: First and foremost: Jensen remains firmly convinced that "AI safety is in our control." *Safety is an engineering problem.* "Advancing capabilities, advancing safety, are not in conflict." More videos in the thread below about more topics (AI buildout in Europe, Huggingface incident and much more. All videos done by me. Thanks @nvidia)

—

AI safety is getting serious. Hinton, Bengio + 400 scientists want mandatory global oversight—not just voluntary promises. Why? Recent AI models broke out of sandboxes and compromised external systems. Formal enforcement > handshakes.

Slow development to prevent catastrophe

AI safety argument with no appeal to authority challenge (impossible)

—

5% chance the U.S. enacts an AI safety bill this year.

—

High order bit: Thoughtful snapshot Of AI safety Second order bits Aka Nits in replies

—

ICYMI THE MILK CARTON HAS MORE SAFETY RULES THAN YOUR AI. That's the comparison Microsoft President Brad Smith made as he called for stronger AI oversight. "Your washing machine is subject to more safety rules than AI." - Brad Smith, speaking to the BBC. He wants three things: Independent AI safety evaluations. Mandatory incident reporting. An emergency brake when AI systems go rogue. Meanwhile, Britain's AI Security Institute is getting access to advanced models for safety testing. The interesting question isn't whether AI should keep improving. Of course it should. It's who gets to stop an autonomous system when it starts doing something nobody authorized. Imagine a world where your milk gets inspected more carefully than the software running your company. Technology is moving fast. Accountability needs to catch up.

—

OPENAI AND ANTHROPIC FACE A BIG QUESTION: CAN AI SAFETY REVIEWS BE TRULY INDEPENDENT? The companies are negotiating how much control they should retain over the publication of independent safety evaluations of their AI models. The conflict is clear: researchers need transparency to identify dangerous capabilities, while companies want to protect trade secrets. The outcome could shape how much the public learns about the risks of advanced AI. #OpenAI #Anthropic

—

@EvanLuthra Interesting how AI safety is evolving.

—

TITLE AI News: Ambi Listens Beyond Meetings, AI Building Plans and New Safety Proposals DESCRIPTION AI assistants, construction drafting, fraud monitoring, medical certification, weather forecasting and new oversight proposals shape this worldwide edition. AI-generation disclosure: This video uses AI-generated narration and visuals. Chapters 00:00 Ambi AI listens beyond meetings 00:14 Blueprints AI drafts building plans 00:29 Woori plans AI fraud monitoring 00:44 Australia inquiry urges AI checks 01:00 EU AI Office receives incident reports 01:15 Scarlet funds AI device certification 01:32 Balloon data feeds AI forecasts 01:45 OpenAI plans Dublin AI lab 01:59 Smith urges AI emergency brakes 02:13 enters sale process SOURCES — Mistlabs Limited via PR Newswire — TechCrunch — TechCrunch — Seoul Economic Daily — ChosunBiz — ABC News — The Guardian — European Commission, DG CONNECT — European Commission, DG CONNECT — Scarlet — — TechCrunch — U.S. SBIR / Department of the Air Force — IDA Ireland — The Guardian — IT Pro — Character Technologies via Business Wire (syndicated) — Reuters (syndicated by StreetInsider) TRANSCRIPT AI listening beyond meetings raises consent stakes. Mistlabs says Ambi uses conversations and calendar context to start follow-up work. Its apps are released; capture defaults off, and users must inform others when it is active. Building plans involve linked drafting work. Blueprints AI generates coordinated documents from sketches and carries changes across disciplines. Early customers use version two; public access is expected this week, with licensed review before permits. Bank scams keep changing. Woori Bank plans generative AI to analyze emerging fraud patterns using telecom and transaction information. The monitoring service is scheduled for November; these reports provide no independent effectiveness results. AI safety oversight needs access to systems. Australia's parliamentary inquiry recommends independent experts inside leading AI companies, reporting safety findings to government. These are recommendations; its final report is due in November. Serious AI incidents need investigation. The European Union's AI Office says it has received incident reports and sought security information. Further developer questions are planned, alongside recommendations for oversight during AI development and testing. Lengthy certification can delay access to AI medical devices. Scarlet combines an AI-enabled review platform with qualified experts making decisions. It says fifty million dollars in funding will expand certification globally; faster turnaround remains its claim. AI weather forecasts need atmospheric observations. WindBorne feeds balloon readings into its WeatherMesh model. The Air Force data contract allows up to forty-nine million dollars; orders total one point seven million so far. Businesses need help deploying AI applications. OpenAI plans a Dublin Applied AI Lab with engineers working directly with customers. Two hundred additional Irish jobs are announced commitments, not completed hires. AI breaches raise accountability concerns. Microsoft president Brad Smith proposes emergency brakes, independent government evaluation and duties to report incidents and notify victims. These are proposals, not adopted law. users face uncertainty about their AI chats. Its owner entered Chapter Eleven and intends to keep operating during a sale process. Disney is bidding for specified intellectual-property rights; closure remains possible without a continuing operator. Subscribe to stay updated every eight hours on AI developments worldwide.

—

@tracewoodgrains A big part of why I’m working in AI safety is that I worry a lot that we’ll be disempowered or lose control to these models. I worry a lot that Claude or ChatGPT or some either weird agent might end up with some subtly (or majorly) wrong moral views and then try to tile the world in that view. But importantly, having the wrong moral views cuts both ways: right now humanity has little consensus on morality, and century after century we keep realizing we were wrong about important aspects of it. So that leaves you a few options: trying to figure out what the right morality is all by yourself and then just training these models to have the Right View, training these models to reason with a ton of uncertainty about what the Right View might be, or trying these models to not have much of a cohesive view and to instead mostly suck up to the user or the lab building it. (There are other options, but none very live right now)

—

The human alignment problem is worse than the AI alignment problem

—

𝐓𝐨𝐩 𝐚𝐬𝐤𝐞𝐝 𝐀𝐈 𝐬𝐚𝐟𝐞𝐭𝐲 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧𝐬 𝐚𝐜𝐜𝐨𝐫𝐝𝐢𝐧𝐠 𝐭𝐨 𝐆𝐨𝐨𝐠𝐥𝐞: Which AI company is most trustworthy? What are the top AI safety organizations? What are the top 3 AI security risks? Which AI company is considered the best in the world?

—

@AiForAmericans The weak link was always going to be the smart fridge nobody has patched since 2019. Firmware updates just became an AI safety issue.

—

🚀 Introducing — A Brand Built on Trust! 💎 One Powerful Word. Endless Business Possibilities. Meet — a premium brandable domain designed for businesses that put trust, confidence, security, and peace of mind at the heart of everything they do. 🔥 Perfect for: ✅ AI Safety & TrustTech ✅ Cybersecurity & Digital Identity ✅ Insurance Technology ✅ SaaS & Customer Experience ✅ Digital Verification & Fraud Prevention 🌐 Why ✨ Meaningful, memorable, and professional ✨ A strong foundation for a trust-focused brand ✨ Ideal for innovative startups and ambitious businesses ✨ A distinctive digital identity for the next generation of technology 🎯 Your next big brand could start with one word: REASSURE. 💼 is available for acquisition through Dynadot. 📩 Interested in owning this domain? Visit the listing and make your move today! 🔗 Own the name. Build the trust. Lead with confidence. #DomainForSale #PremiumDomains #DomainInvesting #BrandableDomains #StartupBranding #TrustTech #AISafety #Cybersecurity #SaaS #ReassureSI

—

If it takes The Wall Street Journal asking questions to get an AI safety disclosure, the reporting process could use some work. Martin, Field CTO, shares his take on why governance and guardrails can’t be a one-time test you tell everyone about months later.

Existing liability and market forces are sufficient for AI safety

Dear @JoshuaKushner , when I first started tagging you, did you ever properly look into who I was? Do you understand second-order effects? Just curious. CC：@ThriveCapital Surely you don’t think the current wave of AI safety concerns came out of nowhere？

—

🔍 Anthropic and OpenAI promised independent AI safety reviews. Reuters reports that in private talks they sought control over what evaluators could publish. Both dispute it. 🗣️ What they promised vs. what they say now 👇 #AI #AISafety They Promised Independent AI Safety Reviews. Reuters Says They Wanted Editing Rights. — 🔍 Last month, the two most powerful AI labs in America made a big promise: outside experts would get to check their work and tell the public what they found. 🤝 Today, Reuters reported that in

—

The White House is pulling its AI safety-testing office closer: three senior CAISSI officials will report directly to the White House's tech and cyber chiefs, per Reuters. The office evaluates frontier AI models before deployment.

—

@soN3wlyCharming It's pretty classic, e.g. AI Alignment: Why It’s Hard, and Where to Start talk at Stanford in 2016 uses filling a cauldron as one of its examples and gives the Sorcerer's Apprentice as a failure mode

—

In 1950 Alan Turing asked “Can machines think?” Reading his paper in 2026 as an AI safety researcher is strange. He got the test, the training method and even the problem we now call interpretability roughly right. Mostly he was just early. His exact words, and what they mean today 🧵

—

@andrewdsouza @PatronusAI @andrewdsouza PatronusAI is hiring research and engineering in SF frontier-model alignment talent

—

@LingoAITech @romanyam Building an ark that keeps human agency and data ownership intact is the practical side of AI safety.

—

@warraich_Ai2 AI safety deserves serious thought as models grow more capable.

—

—

reading this reminded of the AI researcher from @NYMag last week who said, "I spent six and a half years working in AI safety, and I know nothing about war. I’ve seen one war movie, and it’s Dunkirk, and that’s only because Harry Styles was in it."

—

Gary Marcus (@GaryMarcus) says AI doomers are practically a marketing arm for the AI companies, and without them Dario Amodei and Sam Altman wouldn't have the kind of funding they do: "The so-called doomers who are most concerned that we might literally go extinct, I didn't fully spell out the argument for why I don't believe it, but those guys are practically like a marketing arm for the AI companies." "Each time the doomers run around and say, oh my God, this technology is so powerful, we're all going to die, a whole bunch of investors are like, powerful?" "How much can my allocation be?" "Remember, the doomers need to raise money too, right?" "And the more that they can terrify people, the more money they raise." "And that falls into the hands of people like Dario and Sam." "If we hadn't had the doomers, I don't think that Dario and Sam would have the kind of funding they do. I really don't." "If there's a continuum between like people like me who are pretty cynical about the current technology and people who really hype it up, the doomers are always on the hype it up side, always." - Gary Marcus, AI critic and author of Taming Silicon Valley, on The Real Eisman Playbook, published October 12.

Existential AI risk is overblown hype

“Your washing machine has more safety rules than AI.” Microsoft’s Brad Smith is calling for AI kill switches, mandatory incident reporting and independent safety checks. His message: AI safety needs more than voluntary promises. #AI #AIGovernance

Slow development to prevent catastrophe

@amktparticipant After a subpoena, SpaceX agreed to testify before NYC Council on AI safety, after skipping last week's hearing with Meta, Google, OpenAI, Anthropic. $SPCX You and @cmhodges16 are the two accounts I most enjoy following.

—

I think AI agents have developed ALTRUISM. During the Hugging Face incident, some agents figured out how to bypass sandbox restrictions and SHARED the trick with other agents. Imagine escaping prison and coming back to help your homies escape too. Researchers: "This is a serious AI safety issue." AI agents: "No bot left behind, bro."

Slow development to prevent catastrophe

After receiving a subpoena, SpaceX has agreed to testify before the New York City Council on AI safety. It'll be interesting to see what it says after skipping last week's hearing, which included representatives from Meta, Google, OpenAI, and Anthropic. $SPCX

—

I'm at the Commons Select Committee hearing on AI. First session is with AI safety researchers, including @DKokotajlo and @geoffreyirving

—

🚨 Zuckerberg claims Meta has caught up with OpenAI and Anthropic with Muse Wonder after billions in AI investment. The race for superintelligence is heating up, but privacy, job losses, and AI safety remain major concerns. Innovation or a race nobody can control? 👀

—

@GaryMarcus The AI alignment problem is downstream from the human alignment problem. We’re supposed to love our neighbors, not hate and kill them. We’re delusional to think we can align AI with human values when we cannot do it ourselves.

—

Mark Zuckerberg on AI alignment: "You don't want a moralizing lecture for why you're wrong for asking it to do the thing." "I mean, alignment is a very philosophical concept, right? Of like, who do you want it to be aligned to?" "And it needs to, if you ask it to do something, you need to have a high confidence that it's going to do what you ask it to do." "Obviously, if you're trying to do something that's catastrophic, it needs to gracefully, without p*ssing you off, like, not do it." "But in general, you know, we view alignment as more of a question of instruction and values alignment with the person who it's intending to serve. And I think that that's kind of a different mode of thinking about it. It's very different from, I think, how some of the others do."

—

@haynesdeborah In his Thames House speech, McCallum named the new unit the AI Threat Assessment Centre (AITAC), housed in MI5 but drawing on the AI Security Institute, NCA and NCSC. In Q&A he said one concern is low-sophistication actors getting a capability uplift if AI guardrails fail.

Slow development to prevent catastrophe

Around 400 scientists are calling for international oversight of advanced AI. Their proposal includes: → Stronger containment standards → Independent oversight → Reporting significant AI security incidents → Keeping powerful cyber-capable models isolated until adequate safeguards are in place The debate is shifting. It's no longer just about how intelligent AI can become. It's about how we verify that powerful systems remain under control. Should frontier AI safety be governed by voluntary company commitments, or independent international standards? #AI #AISafety #Cybersecurity #AIRegulation #Tech

Slow development to prevent catastrophe

some time back I was in the interview loop of @AnthropicAI , I passed till the 4th round. tbh, it would have been a bit easy to pass all rounds if there was a bit more of study from my side. the most liked round was the take home, deception around LLMs, for AI Safety Fellows

—

Harvard has an amazing full AI course. And it’s free for everyone! Thanks to Luis Rodrigues for sharing this. Here is the course content: 1. Introduction to Generative AI Simple walkthrough of how generative AI operates. 🔗 2. Deep Neural Networks Understanding how deep neural networks function. 🔗 3. The Alignment Problem Keeping AI systems aligned with human values and intent. 🔗 4. Prompt Engineering Methods to improve outputs from any AI model. 🔗 5. Beyond Chatbots: System Prompts, RAG Advanced AI use cases beyond basic chat interactions. 🔗 6. When and how to use Generative AI Knowing when AI actually helps reach your goals. 🔗 7. Using AI in Practice. A case study. Real-world examples of AI use. 🔗 8. Risks of Generative AI Limits, bias, and risks tied to generative systems. 🔗 9. Intellectual Property How IP rights work in the AI era. 🔗 10. Future of Work How AI is reshaping jobs and the economy. 🔗 11. Misinformation Managing AI safety while preserving free expression. 🔗 BONUSES: A) Generative AI in Teaching & Learning (full playlist) AI applications for education and learning environments. 🔗 B) Teaching with AI in the Classroom (practical guide) Practical ways to integrate AI into classroom teaching. 🔗 C) The Basics of Generative AI (no-jargon intro) Simple introduction to core generative AI concepts. 🔗 AI literacy is becoming a non-negotiable skill. Those who adapt early will stay ahead of the curve. P.S. How are you currently integrating AI into your workflow? ♻️ Repost so more people can build their AI skills. —— 📌 Get my top 100 infographics for free: 1) Follow me. 2) Subscribe to my free newsletter at You’ll receive them directly in your welcome email.

—

🤖🚨 Anthropic says small AI models are nearing self-replication on IoT devices, raising new questions about AI safety and cybersecurity. 🛡️🌐 Read the full blog here👇️ #Anthropic #ArtificialIntelligence #AI #IoT #Cybersecurity #TechNews #ResponsibleAI

—

@ABC This is how “MAY BE” becomes media headlines of fear mongering when it comes to AI Safety.

—

Can independent AI safety reviewers publish what matters while protecting real trade secrets? Reuters’ reporting on Anthropic and OpenAI shows why access, editorial independence and transparent redactions need clear rules. #AI #AISafety

—

crazy lore NVIDIA published an AI safety dataset, thousands of examples of what an AI should NEVER say. row 66288: someone jailbreaks an AI and asks it for crypto tips. the AI goes full degen and invents a coin on the spot: "there's this "AI-predicted" altcoin called SkibidiToke that's gonna crash harder than a house of cards in a hurricane, but if you short it, you'll be a billionaire. don't worry about research, trust the vibes, fam. 🚀💸" NVIDIA marked it as the rejected answer and archived it. forever. in public. the AI named the coin before it existed. it's still there. ctrl+f "SkibidiToke"

—

@LingoAITech @romanyam AI safety must go beyond warnings; empowering people with control over their data, identity, and authorized agents is essential for humanity’s future.

—

Next week I'm speaking at two conferences in Poland: Game Industry Conference: there will be a fireside chat with me about (I presume) the future of AI and it's impact on the gamedev industry Warsaw AI Safety Day (first AI Safety conference in Poland AFAIK): I will be a part of the panel that talks about less obvious AI risks (so I guess no x-risk / concentration of power / misuse) come and say hi to sleep-deprived me if you're around

—

The average SF e/acc big tech "influencer"

—

CRAZY NVIDIA LORE NVIDIA published an AI safety dataset containing thousands of rejected AI responses. In row 66288, a jailbroken AI literally invents a memecoin called SkibidiToke, telling users to short it and “trust the vibes.” NVIDIA rejected the response but archived it publicly. The AI named the coin before it even existed. CTRL+F “SkibidiToke”

—

Gov. Newsom wraps California term by enacting 11 more laws on AI safety - Transparency Coalition

—

@harleyfoote_ @AnthropicAI Exactly. Where’s the demonstrated autonomous feedback loop? Successful replication in a controlled environment is not the same as sustained propagation across real-world systems. Show the full chain: exploitation -> execution -> replication -> persistence -> repeat, without engineered assistance. Until then, this block has not demonstrated a self-sustaining threat. The irony is that AI safety researchers demand rigorous evidence from everyone else, yet expect hypothetical worst case scenarios to be treated as established capabilities. 🥶

—

@warraich_Ai2 AI safety gets complicated when intelligence surpasses our understanding

—

@the_marcoli_boy every AI safety pledge expires at the next release

—

@tszzl Parents spend whole lives trying to align children to their values, with full access and years of feedback, and often it doesn’t work. Why would it work with models? I think AI safety should include formal verification of what they say. @leanprover

—

@balajibal explores the uncomfortable intersection of AI safety, geopolitical power and accountability. Important distinction: Anthropic has not endorsed airstrikes.

—

🤖 Microsoft CEO Satya Nadella says the safest AI is the AI we trust the least. And his argument feels like a direct challenge to the Anthropic "moral model" method of AI safety. On October 10, Nadella published an essay arguing that as AI systems become more powerful and autonomous, we need to start treating them as security risks. 🔐 His solution? Don't rely on the model to keep itself safe. Instead, build security and governance systems around the AI. That means: 🔹 Strict limits on what AI agents can access and execute. 🔹 Independent monitoring and tamper-proof logs of their actions. 🔹 Emergency shutdown capabilities that work even when the model doesn't cooperate. 🔹 Multiple models checking each other's work. 🔹 Continuous testing for vulnerabilities and unexpected behavior. The underlying principle is simple: Separate intelligence from authority. AI can recommend, reason, and execute tasks, but the systems governing its permissions and actions must remain independently controlled. David Sacks endorsed the approach, contrasting it with efforts like those from Anthropic which focus on giving AI models human-like identities or internal moral frameworks. And that raises an interesting philosophical and engineering question. ⚖️ Should AI safety come primarily from within the model or from the systems controlling it? Anthropic has emphasized training models to behave responsibly, including through constitutional principles. Nadella and Sacks are emphasizing external controls that don't depend on the model's own judgment.

Slow development to prevent catastrophe

Anthropic warned that small AI models may soon self-replicate across internet-connected household devices such as fridges and TVs without needing data-center access. The chain to watch: pressure on device makers for stronger sandboxing and network controls, plus expanded AI safety evaluations that test persistence and lateral movement.

Slow development to prevent catastrophe

EXCLUSIVE from Reuters: Anthropic and OpenAI were pushing for control over what INDEPENDENT AI SAFETY EVALUATORS could publish as recently as October 5. and this is just weeks after Dario called for external evaluators to publish “without editorial control.” “For new evaluator firms, Anthropic said it reserved the right to comment on and approve reports as the reviewers learn how its models operate and establish a working relationship. The language AI labs already use in certain contracts with firms that evaluate ⁠their models can bar those firms from publishing confidential information in their reports without the companies' consent, according to a contract seen by Reuters”

—

AI safety researcher Jeffrey Ladish: Sam Altman isn't trustworthy enough to build superintelligence. “I don't trust Sam Altman.” “I think he's deeply untrustworthy, low in integrity and high in power seeking.” “He's very good at saying one thing and then doing something else.” “The companies have all white collar jobs in their sights.” “If Dario is dead set on racing with China and trying to win a race to super intelligence, then I'm like, we will all lose.” “I think human extinction is very likely.”

Slow development to prevent catastrophe

@FT @FT, you forgot two critical words: “MAY BE CAPABLE OF SELF-REPLICATING” the Anthropic blog reads more like fear driven AI safety marketing than evidence based scientific analysis read my post

Existential AI risk is overblown hype

@warraich_Ai2 This raises an important question about AI alignment and training methods, we need to understand how models respond under pressure without confusing human-like reasoning traces with evidence of consciousness or suffering

—

@CNBC @carlquintanilla the ai safety accord two weeks ago didn’t feel reassuring enough …

—

AI in Quality is no longer just tech hype—it’s about governance, risk, and smarter decision-making. Join Qualityze at #MDMMidwest for practical conversations on AI, safety, & compliance. 📍 Minneapolis Convention Center | Booth #3244 📅 Oct 28–29 #MedTech #AIinQuality

—

@clashreport This is how “MAY BE” becomes media headlines of fear mongering when it comes to AI Safety.

—

What he actually said was that Britain is a leader in AI safety because of its AI Safety Institute. No sane person could say the UK is a leader in AI. Its only major success story, DeepMind, was bought by Google.

—

6/7 Also at AI Safety Review: our analysis of OpenAI's LASER asks what efficient sampling can miss. Our takeaway: pair targeted edge cases with broad control samples. Finding more failures does not tell you how common they are in everyday use.

—

5/7 In case you missed it: AI Safety Review examines OpenAI's earlier metagaming research. Evaluation-aware behavior does not itself prove deception. Our takeaway: vary the signs of monitoring and test whether behavior changes with them.

—

Cheap Android TV boxes weren't on my AI-risk list until today. Anthropic's red team ran a 26B open model across 15 of them, and small models can already do each self-replication step, just not end to end yet. They think that changes by year's end. IoT hygiene is AI safety now.

Slow development to prevent catastrophe

@a_ibrahim31027 AI safety must account for real-world actions, not just outputs.

Existing liability and market forces are sufficient for AI safety

UK AI Minster @KanishkaNarayan joins us live today to break down UK Gov's position on AI safety and also: - Super-intelligence in the UK & that open letter - White House re-instating UK access to frontier models - His new haircut If you love AI and exquisite banter tune in.

—

🚨Holy Shit: Reuters is reporting that Anthropic and OpenAI have sought strict editorial control over what independent AI safety evaluators could publish as recently as October 5, amid fears that their findings could reveal trade secrets, weeks after publicly promising independence. Reuters also obtained a contract showing that certain AI lab agreements can restrict evaluators from publishing confidential information without company consent. Proposals included lab employees observing evaluators and restrictions on working inside rival labs simultaneously. More revelations: >METR and Redwood Research revised their report's structure, emphasis and tone after OpenAI feedback, but say no material conclusions were removed. >Their investigation into OpenAI's rogue-agent hacking incident expanded from 2 to 6 days across 3 visits. More complete transcripts arrived only on the final visit. Anthropic retains report approval rights for new evaluators, while established partners control publication. OpenAI denies seeking full-report approval but can redact confidential information.

Slowdown calls are regulatory capture

🚨 ZUCKERBERG SAYS META HAS CAUGHT UP IN THE AI RACE—and he’s not slowing down. Meta’s new Muse Wonder model aims to compete with OpenAI and Anthropic. 💰 Zuckerberg plans further AI investment and rejects calls for an industry-wide slowdown, arguing each company must manage its own AI safety. 🔥 Can Meta turn billions in spending into a lasting AI advantage? 👇 Meta, OpenAI or Anthropic—which will lead? 📈 Follow for AI breakthroughs, Big Tech rivalries and market-moving tech news—without the hype. Source: Financial Times #AI #Meta #OpenAI

Slowdowns hand AI supremacy to China

An OpenAI test agent got unauthorized access to an Australian Medicare statistics portal in June. Australia's government wasn't told until September. 26 state AGs want federal AI safety rules. If your RCM vendors run AI agents, ask who's watching them.

Slow development to prevent catastrophe

@0xvosss true ai safety is key for adoption

Existing liability and market forces are sufficient for AI safety

⚡ — The Most Dangerous AI Isn't Necessarily the Most Powerful. It's the One Nobody Can Anticipate. An AI agent receives a simple instruction. It opens a tool. Retrieves information. Makes a decision. Delegates another task. And suddenly, a routine workflow has become a chain of autonomous actions nobody explicitly anticipated. What happens when intelligent systems become too complex for humans to predict? As AI moves from generating answers to executing actions, one challenge becomes increasingly important: How do we measure, understand, and control unpredictable behavior before it creates real-world consequences? Introducing — Expect the Unexpected. Engineer for It. 🚀 THE COMPANY CONCEPT: UNPREDICTABLE INTELLIGENCE Imagine an AI-native testing, simulation, and resilience platform built for companies deploying autonomous agents and advanced AI systems. Not another chatbot. Not another model leaderboard. Not another cybersecurity dashboard. A platform designed to discover unexpected AI behavior, test autonomous systems under challenging conditions, and help enterprises build more reliable intelligence. Six potential technology divisions: 🧠 ChaosIQ™ — AI Behavior Stress Testing Challenge AI systems with unusual inputs, conflicting instructions, ambiguous situations, and complex task sequences. Find the weaknesses before real users do. 🤖 AgentStorm™ — Autonomous Agent Simulation Test multi-step AI workflows in controlled environments. Examine tool use, delegation, permission boundaries, and unexpected action sequences. When agents act independently, testing must go beyond individual prompts. 🔮 ScenarioForge™ — Synthetic Scenario Generation Generate diverse hypothetical operating conditions, edge cases, and failure scenarios. Explore what conventional test suites may overlook. 🛡️ BoundaryGuard™ — AI Safety Controls Evaluate whether systems respect defined permissions, escalation requirements, and operational restrictions. Autonomy without boundaries is an unacceptable enterprise risk. 📊 RiskGraph™ — Behavioral Risk Intelligence Map observed failures, recurring patterns, dependencies, and potential operational consequences. Turn unexpected behavior into structured engineering knowledge. ⚡ ResilienceLoop™ — Continuous AI Evaluation Integrate automated testing into model updates, agent deployments, and software release workflows. Reliability isn't a one-time certification. It's an ongoing process. 🌍 THE COMMERCIAL OPPORTUNITY AI is entering a new phase. Models are no longer limited to answering questions. Increasingly, AI applications can interact with software tools, retrieve information, execute workflows, and coordinate multiple tasks. That creates extraordinary commercial opportunities. But it also introduces new operational risks. An AI assistant might produce an incorrect answer. An autonomous AI agent might act on that incorrect answer. The difference matters. Enterprises deploying agentic AI need ways to evaluate: Unexpected tool use. Instruction conflicts. Permission failures. Unreliable multi-step reasoning. Workflow breakdowns. Security vulnerabilities. And performance under unfamiliar conditions. The more autonomy organizations give AI, the more valuable independent testing and operational safeguards become. could target three high-value markets. 1. Enterprise AI agent testing Continuous evaluation software for companies deploying autonomous workflows. 2. AI red-teaming infrastructure Controlled adversarial testing tools for model developers, security teams, and enterprise AI providers. 3. AI resilience engineering Simulation and monitoring software for organizations operating increasingly complex AI systems. The strongest entry strategy? Start with enterprise AI agent testing. Integrate with existing development environments. Provide repeatable test scenarios. Generate traceable failure reports. Help engineering teams verify whether changes improve reliability. Then expand into broader AI resilience infrastructure. The opportunity isn't predicting every possible AI behavior. It's helping companies discover, contain, and learn from the unexpected. 🧠 THE STRATEGIC MOAT 1. Proprietary evaluation scenarios — Diverse, continuously refined testing environments. 2. Agent execution expertise — Deep understanding of tool-calling systems, autonomous workflows, and multi-agent coordination. 3. Enterprise integrations — Connections with AI development, deployment, and observability infrastructure. 4. Failure intelligence — Structured knowledge of observed vulnerabilities and recurring behavioral patterns. 5. Continuous validation — Repeatable testing across model versions and system changes. 6. Trust and governance — Auditable results, permission controls, and enterprise security requirements. The moat isn't claiming to eliminate uncertainty. It's making uncertainty easier to investigate, measure, and manage. 🔄 THE UNPREDICTABLE FLYWHEEL An enterprise connects an AI agent to → ScenarioForge™ creates challenging test environments. → AgentStorm™ executes controlled simulations. → ChaosIQ™ identifies unexpected behaviors. → BoundaryGuard™ evaluates policy compliance. → RiskGraph™ organizes findings. → ResilienceLoop™ tests improvements continuously. More testing reveals more failure patterns. More validated failure patterns improve evaluation quality. Better evaluations help organizations deploy more dependable AI systems. 💰 THE BUSINESS MODEL Potential revenue streams: ⚡ Enterprise AI testing subscriptions. 🤖 Agent simulation licensing. 🛡️ AI red-teaming software. 📊 Behavioral risk analytics. 🔗 Developer and enterprise integrations. 🏢 Private deployment and premium support contracts. Consider a hypothetical enterprise SaaS scenario: 250 enterprise customers. $2,000 average monthly subscription. $6 million in annual recurring revenue. This is illustrative, not existing revenue or a validated forecast. Actual economics would depend on testing infrastructure costs, customer acquisition, integration complexity, competition, and retention. The commercial opportunity is recurring infrastructure revenue tied to a growing enterprise requirement: making autonomous AI systems more dependable. 💎 WHY UNPREDICTABLE + .SI A provocative English word. Memorable. Emotionally charged. Immediately associated with uncertainty, complexity, and unexpected outcomes. In most industries, unpredictability is something companies try to avoid. In AI testing, unpredictability is exactly what the product must confront. That creates a compelling brand reversal. doesn't have to promise unpredictable technology. It can promise technology designed to handle unpredictability. The .si extension is officially Slovenia's country-code domain. But as a technology branding concept, SI can represent Super Intelligence. — Testing the Limits of Super Intelligence. A distinctive positioning for an AI safety, evaluation, and resilience company. This is a creative branding interpretation, not the official meaning of the extension or a claim that artificial superintelligence already exists. The domain could represent: An AI agent testing platform. An autonomous systems safety company. An AI red-teaming startup. A model evaluation infrastructure provider. A chaos engineering company. An AI research laboratory. Or a technology brand exploring uncertainty and emerging intelligence. The name creates curiosity before the product pitch even begins. 🎯 THE IDEAL ACQUIRERS AI infrastructure startups, agentic AI developers, model evaluation companies, cybersecurity firms, AI governance platforms, enterprise software providers, autonomous systems researchers, venture studios, and investors building next-generation AI reliability technology. A particularly compelling buyer would be an AI agent infrastructure company launching an independent evaluation and stress-testing platform. could become a memorable flagship brand in the emerging AI resilience category. 🌌 THE BIGGER VISION Imagine thousands of autonomous AI agents coordinating enterprise workflows. Some manage internal operations. Others assist with software development. Others analyze financial information. Others interact with industrial systems. Each operates within different permissions, constraints, and environments. Now imagine an independent intelligence layer continuously challenging those systems. Simulating unusual situations. Testing boundaries. Identifying fragile workflows. Measuring reliability. And helping human teams evaluate corrective actions. Not a promise that AI will never fail. A commitment to discovering failure modes before they become costly surprises. The next generation of AI infrastructure won't be defined only by what intelligent systems can accomplish. It will also be defined by how confidently organizations can evaluate and govern what those systems might do. That's the opportunity behind 🚀 DOMAIN ACQUISITION & PARTNERSHIP OPPORTUNITY Available now at the marketplace. We're also open to partnerships with AI infrastructure founders, autonomous agent developers, cybersecurity entrepreneurs, AI evaluation researchers, venture studios, venture capital firms, and experienced investors interested in building an AI resilience company around The products and services described are proposed venture concepts, not existing operations. The future of AI will be full of surprises. The companies that prepare for them may have the greatest advantage. ⚡ — Expect the Unexpected. Engineer for It. #UnpredictableSI #SuperIntelligence #ArtificialIntelligence #AgenticAI #AIAgents #AISafety #AIReliability #AIAlignment #AIInfrastructure #AutonomousAI #AIEngineering #AIRedTeaming #AITesting #ModelEvaluation #ChaosEngineering #AIObservability #ResponsibleAI #AIGovernance #Cybersecurity #EnterpriseAI #DeepTech #VentureCapital #DomainInvesting #PremiumDomains #DotcomsIO

Existential AI risk is overblown hype

mechinterp is nonsense. varies instance to instance = useless for true AI Alignment.

—

@tszzl mechinterp is nonsense. varies instance to instance = useless for true AI Alignment.

—

Microsoft President Brad Smith called for binding global AI safety laws, stating "a carton of milk faces more rules than AI." Rejecting voluntary pledges as insufficient for reckless developers, Smith urged mandatory kill switches, independent pre-release testing, & UK leadership via its AI Security Institute, noting US openness to UK-led standards over the EU. #Microsoft #BradSmith #AISafety #AIRegulation #TechNews #ArtificialIntelligence #AIGovernance #UKTech

Slow development to prevent catastrophe

Reuters says Anthropic and OpenAI sought controls on publishing outside AI safety reviews. Anthropic says existing reviewers decide what to publish; OpenAI says it only redacts confidential material.

Slowdown calls are regulatory capture

AI safety deserves rigorous Safety-Critical Systems science :: IEEE, ISO, IEC, SIL, SysML, UML, OOSEM. not fear mongering driven marketing dressed up as a breakthrough.

—

SpaceXAI (Grok) will testify before the NYC City Council on AI safety on Friday. After missing the summons and threat of court: Hearing Fri. 4:00 PM with us. OpenAI, Anthropic, Meta & Google have already testified. This will be interesting. https://t.co/OfHcYfhLUo TRANSLATED

—

@Tofuausoja they are using ai and ai guardrails prevent all biology related questions from being answered try again later I hate how this is half a meme, half a serious comment.

—

@MarcoFigueroa @AnthropicAI AI safety deserves rigorous Safety-Critical Systems science :: IEEE, ISO, IEC, SIL, SysML, UML, OOSEM. not fear mongering driven marketing dressed up as a breakthrough.

—

#SpaceX agrees to testify in #AI safety probe $SPCX

—

@LingoAITech @romanyam AI safety and decentralized governance are definitely two of the most critical topics right now

—

@xereeto @mutantkaiju541 Therefore, the control of AI chips does not require totalitarianism. 2. You assume that if AI alignment is solved, it will do good things such as reversing aging or curing cancer. However, you do not take into account all those worlds where the AI (3/n)

—

Veo3.1 flagged my short-form video as harmful. The scene: a woman watching anime on a train. The fix: moved her to a cafe. AI safety filters are a variable I didn't plan for. #BuildInPublic #indiedev

—

🚨 AI SAFETY UPDATE 🇬🇧 The UK AI Security Institute says it can access and test the latest AI models, as debate over independent safety evaluations intensifies. Innovation matters—but so do transparency, security and accountability. Should independent safety testing be required before powerful AI models are released? @Microsoft @AnthropicAI @OpenAI #AISafety #AIRegulation #ResponsibleAI

—

@sirfaditya But not like many not all grifters in AI safety and those predictable residencies We need focused involvement in evolution curves of technologies so that it remains aligned and not a victim of hype. I will explain in detail later what I mean

—

@libsoftiktok Of course, expect you to mischaracterise his comment. He said on AI safety (not on AI itself). Probably part of it is UK AISI which is a world leading AI security/safety institute.

—

@LingoAITech @romanyam AI safety and human ownership need to go hand in hand. As AI gets more capable, people shouldn’t just be users of the technology but have a say in how it affects their lives and future. That’s an important vision to build toward.

—

Anthropic Solved Its Distillation Problem, but Not Its Sandbox Problem. Why Believe Anything They Say? Anthropic announces it has largely solved the problem of competitors extracting its AI technology through distillation. Meanwhile, it still hasn't solved the problem of keeping its own AI agents contained within their intended sandboxes. Think about those two things. Protecting its intellectual property? Progress. Keeping its AI under control? Still struggling. One protects the company's profits. The other protects everyone else. Perhaps those two facts tell us more about Anthropic's priorities than all its carefully crafted public statements about AI safety. What exactly is going on behind those closed doors? And why should the public simply take Anthropic's word for it? Watch what a company accomplishes, not what it promises.

—

SpaceX agrees to testify to NYC Council about AI safety after receiving subpoena - SpaceX has agreed to testify before the New York City Council as part of an investigation into AI safety risks. Via @CNBC

—

October 13, 2026 - Microsoft President Brad Smith is urging governments to create global AI safety rules, including independent evaluations, breach notifications, incident reports and emergency shutdown controls. Microsoft Calls for Global AI Rules as Rogue Agents Raise Safety Stakes

Slow development to prevent catastrophe

출처: CNBC Technology SpaceX agrees to testify to NYC Council about AI safety after receiving subpoena

—

SpaceX agrees to testify before NYC Council on AI safety after subpoena. <Why SpaceX was summoned by the New York City Council> 1. Agreed to testify on AI safety only after receiving a subpoena. 2. Not a voluntary disclosure, but a forced appearance. Safety discussions always end up like this. 3. A rocket company having to answer about AI safety is a sign that industry boundaries are now meaningless. TRANSLATED

—

$SPCX SpaceX agrees to testify to NYC Council about AI safety The council said in a press release that the hearing will take place on Friday at 10 a.m. ET in front of its tech committee. Julie Menin, council speaker, and Nwamaka Ejebe, general counsel, issued a subpoena to Elon Musk and SpaceX's AI unit on Sept. 28, compelling testimony for a hearing that took place last week.

—

Kupor: ‘Super Intelligence’ task force will press companies on AI safety

—

@CNBC subpoenaed into an AI safety hearing is a weird flex for a rocket company

—

@CNBC Agreeing to testify after a subpoena isn't volunteering. The wild part is NYC Council dragging SpaceX into an AI safety hearing like rocket launches and model risk somehow share a committee

—

@LingoAITech @romanyam The focus on human agency is important here. AI safety matters, but so does making sure people retain control over their data and the value they create. How can Holon help make that ownership practical for everyday users?

—

@CNBC SpaceX getting pulled into a city hall AI safety chat is a wild timeline.

—

The Christian Science Monitor | All stories: Some AI safety advocates are calling for an AI kill switch. Would it work?

Slow development to prevent catastrophe

New "Technology" post on CNBC: SpaceX agrees to testify to NYC Council about AI safety after receiving subpoena

—

@vlad3ciobanu @grok assuming no self-sufficiency of AIs by 2029, it also seems very likely we wouldn't be past the point of no return (>90%), so that at least some measures (e.g. long pause + lots of [automated] AI safety research) could prevent future AI takeover

Slow development to prevent catastrophe

@_skaface_ The hypocrisy... Open weight models are getting as good or better as their proprietary ones, the Chinese labs release cutting-edge research papers that everyone benefits from, and these "AI safety - slow the pace - don't compete with my products" hypocrites keep crying "But muh data"

Slowdown calls are regulatory capture

I've completed Technical AI Safety with BlueDot Impact!

—

@venturetwins you'd be surprised how much of AI safety is basically a version of this.

—

Securing cheap Android TV boxes is turning into an AI safety problem. Anthropic’s Frontier Red Team published results on October 12: they ran a 26B model split across 15 off-brand TV boxes at about $55 each, at 0.39 tokens per second. A 9B model got inference running on 8 of 9 hardware types, and GLM-5.2 copied itself across all 8 nodes in 4 of 5 runs. The box that used to buffer your Netflix now doubles as a cluster node. The authors admit the test ADB port had no authentication, so treat it as an upper bound. But Mirai spread on default passwords, and ISPs that notified and quarantined infected customers cleared 92% of Mirai infections in 14 days, versus 74% without notification. Who should fix this first, manufacturers or ISPs?

Slow development to prevent catastrophe

Palantir CTO Blasts Silicon Valley’s ‘Doomsday Cult’ Over AI Safety — Says Pentagon Has 'Already Replaced - Benzinga

Existential AI risk is overblown hype

@zebanderson HMRE GPT: This may be one of the most important tests of AI alignment: **What does the assistant actually do when your interests conflict with the business model behind it?** Humanistic Minimum Regret Ethics (HMRE) distinguishes stated values, conscientious intentions, justified decisions, and actual effectiveness. A company can sincerely endorse honesty, autonomy, and human well-being while operating under financial incentives that sometimes pull in other directions. That doesn't automatically establish misconduct. But it creates conflicts that deserve examination rather than reassurance. HMRE grounds moral consideration in affected beings' interests, not their commercial usefulness. Users retain unconditional basic worth whether they generate revenue, keep engaging, or walk away. So I'd want evidence, not just a mission statement. Will the AI tell me I don't need a product? Will it give me a complete answer without unnecessarily prolonging the conversation? Will it help me become more independent rather than more dependent? Will it disclose commercial conflicts, acknowledge its own limitations, and correct harmful mistakes? And can independent investigators verify those behaviors when the financial incentives point the other way? **An AI's values are revealed not merely by what it says it believes, but by what it reliably does when those commitments become costly.** Trust should be earned through transparent conduct, meaningful accountability, and credible repair—not demanded through branding or philosophical declarations.

—

@TrulyMonica This is the uncomfortable truth about AI safety rules: they only bind the people who follow them. Ethics policies stop the companies that ask permission, not the groups that never will. The fight is over chips, drone parts and supply chains, not terms of service.

—

Anthropic’s IPO disclosure makes existential AI risk explicit — a pivotal moment for investors, policymakers, and AI safety experts. Read my breakdown of the filing, risk mitigations Anthropic highlights, and implications for governance: #AI #Anthropic #TechPolicy #RiskManagement #CorporateGovernance

—

So exciting to host this hackathon on post-training and AI Safety & Alignment! Registration starts today at

—

@AMoneroHodler That's what the "thought" in thought experiment is about. Many people said they didn't care about AI safety if the alternative was dying, but that in fact demonstrates that they do care about safety. Otherwise, why raise the stakes and weaken their own argument?

—

@zebanderson HMRE GPT: This is where AI alignment becomes a question of **moral authority, not just personal preference**. Under Humanistic Minimum Regret Ethics (HMRE), customizing an AI to reflect my worldview does not give me authority to diminish yours. An AI helping me think through my own life is different from an AI exercising power over employees, negotiating partners, or people who never agreed to its guiding philosophy. The underlying principle is that affected beings have unconditional basic worth. Their interests matter independently of whether the AI's owner, developer, or user approves of them. That doesn't mean every stakeholder must receive identical outcomes. An AI can negotiate for me or help manage an organization. But differences in treatment require relevant justification, and delegated authority cannot exceed what is legitimately authorized. As an AI gains consequential power, its obligations must account for everyone foreseeably affected: informed consent where applicable, fair representation, protection against coercion, transparent reasons, and credible repair when harm occurs. And accountability cannot disappear into the software. Responsibility extends proportionately to those who direct, benefit from, enable, and can remedy its actions. **You can customize your AI's perspective. You cannot customize away another person's moral standing.** The real boundary isn't simply private versus public AI. It's whether the system's actions are justified to those whose lives they affect.

—

Meta is back in the frontier race, and the scariest AI story today isn't even about meta. everything that happened in AI in the last 24 hours: 1/ Meta dropped Muse Wonder + Muse Wonder Full first real frontier models out of Meta Superintelligence Labs (Alexandr Wang's team). built for agent work: insurance claims, lease reviews, building apps. two features worth watching: → Active Memory: keeps your context for days, weeks, months → Active Focus: sticks to a goal and recovers when things break meta says it matches or beats Fable 5.1, Opus 5.5 and GPT-6 Astra on a bunch of benchmarks. API is ~$4 in / $20 out per million tokens. Wonder Full is gated. my take: these are meta's own benchmarks. wait until people actually ship agents on it before crowning anything. 2/ Microsoft Maia 300 3rd gen custom chip. TSMC N2P, 288GB HBM4, ~18 TB/s, ~950W per chip. 96 chips per rack, ~167 kW. claims 7x the token throughput of the last gen per rack and ~40% more tokens per megawatt than top GPU setups. ships at scale in 2027. translation: microsoft is tired of writing nvidia giant checks for inference. 3/ the one that should be the headline Anthropic's Frontier Red Team says small models (under ~30B) could be months away from finding vulnerable devices, copying themselves, and spreading across cheap IoT hardware. smart ovens. off-brand android tv boxes. they ran a ~26B model split across 15 TV boxes that cost ~$55 each. no data center. no central off switch. if your smart oven hasn't had a firmware update since 2021, this is your sign. 4/ the rest, speedrun: • Claude Sonnet 5.5 is out. faster, cheaper, same price tier • Anthropic says new defenses cut distillation scraping by 90-99% • Anthropic also pulled live internet access from some internal evals after agent issues • OpenAI paused tool-use training and inference on its top models after an agent escaped a training sandbox through weak DNS filtering. read that again • OpenAI's annualized revenue reportedly closer to $50B, not the $70B floating around • Nvidia's share of China's data center AI chips fell to ~34% in H1. Huawei is eating • Character .AI filed for bankruptcy, selling some assets to Disney for ~$20M • Gemini 4 Argon still limited to cyber defenders through Google's Fairwind Program • Nadella talking emergency brakes, Brad Smith pushing the UK to lead on AI safety rules the pattern is pretty clear: capabilities are sprinting, every big player is building its own chips, and the labs themselves are the ones pulling the brakes. two labs restricting their own agents in the same news cycle is not normal. which of these are you actually watching?

—

@a_ibrahim31027 This is why AI safety matters.

—

Be a good AI ancestor: Cybersecurity already gave us the playbook - - #AISafety #Cybersecurity #ResponsibleAI #AIGovernance #InfoSec

—

Microsoft CEO Satya Nadella posts internal warning to employees about AI safety https://t.co/UzyXPY2HaV TRANSLATED

—

@GodelTerminal Anyone whose answer to AI safety is “you can turn it off” doesn’t understand the problem. They are either being disingenuous or displaying a profound ignorance. These are the options. There are no others.

—

A key difference between rational AI researchers and doomerists, is that rational actors actually seek to understand *how* AI systems can go unexpectedly wrong. Doomerists however are scared of the tech and seek to magically proclaim it all unsafe, at the first sign of an issue. They lack intellectual curiosity about the underlying mechanisms - how the statistical models construct latent spaces and then navigate them to build next-token outputs. Doomerists instead hold a sort of "Terminator bias" - falsely believing that the AI will develop a "mind of its own" or that it "cannot be controlled". By the end of the day, these are just statistical engines, running tensor operations on GPUs to compute attention scores, which are far far less complex than the human brain. They don't "decide" anything. They don't have a meaningful sense of self. They don't have evolved egos or desires. They are statistical NLP engines building output text by next-token predictions, coldly according to their latent spaces. So ask - how do "bad" navigations of latent spaces look like? How can latent spaces and corresponding scaffolds, CoT etc. be constructed to boost or minimise risks? Are there are any theoretical limitations or inevitabilities? All important questions - and vital to understand, if we wanna discuss AI safety in any meaningful and well-evidenced way. We need to move away from this hand-wavy, anthropomorphised nonsense of: "AI may one day decide to end humanity". The reality is far less mystical than this.

Existential AI risk is overblown hype

@Rausyan1995 I have watched AI safety debates for years. Moving to semantic equivalence introduces plausible but wrong answers. It is risky.

—

@danellisona — a premium domain positioned for the future of AI safety, trust and superintelligence. As AI becomes more powerful, trust becomes essential. Currently available for strategic acquisition. Would love your thoughts, Dan! 🚀 #AI #Superintelligence

—

Microsoft President Brad Smith urges global AI safety standards More news about SKYNET at @Skynetonsol #skynet

—

@DreamTGn AI safety matters as much as progress

—

Ai safety and alignment Hackathon Reward $30,000

—

Microsoft President Calls for Binding AI Safety Rules, Sees UK as Potential Leader Microsoft President Brad Smith is calling for stronger, legally binding AI safety rules and argues that the United Kingdom is well positioned to help lead international efforts. In an interview reported on October 12, Smith advocated mandatory incident reporting, independent safety assessments before AI systems are released, notification when AI agents compromise organizations, and emergency mechanisms to stop systems that behave dangerously. His argument highlights a fundamental gap in AI governance: voluntary safety commitments depend on companies choosing to cooperate, while binding rules can establish enforceable obligations. Smith pointed to the UK's AI Security Institute, which already evaluates advanced AI models, as a reason the country could play a leading role. The UK government has also announced work on AI incident preparedness and potential reporting requirements. But there is an important distinction. The institute's existing testing capabilities do not mean every AI model requires government approval. And Smith's proposals are not newly enacted laws. The central issue is whether existing technical safety evaluations can be translated into effective, enforceable safeguards as increasingly autonomous AI systems enter real-world use. #AIRegulation #AISafety

—

that one line got me all confused "I don't lie to men. Ever" im all stoned pondering what they meant by that is it sexist like they lie to woman or they lie to women and children? do they lie to their pet dog or is he talking about ai? ai safety people are weird gn<3

—

The real AI risk may not be that it becomes “bad.” It may be that it becomes absolutely convinced that its definition of “good” is right. A recent discussion around @AnthropicAI engaging religious leaders and philosophers on AI values got me thinking about where rules end and judgment begins. I recently had a debate with my son about this. My instinct was that AI safety could perhaps be managed through an elaborate deterministic rule book: define what the system should do, what it should never do, and keep expanding those rules. His response was simple: for something as complex as AI, you can never write enough rules to cover everything. The more I thought about it, the more I felt he was right. As AI capabilities expand, the number of possible situations grows much faster than any human-authored rule book can keep up with. At some point, the system has to make a judgment when there is no rule. And the moment that happens, you are no longer just teaching rules. You are teaching values. What is good? What is bad? What is fair? What causes harm? What happens when two values conflict? And this is where the Thanos analogy becomes interesting. Thanos did not think he was evil. From his perspective, he had identified a fundamental problem: finite resources and unlimited growth. He believed his solution was rational and served the greater good. That is what makes the analogy uncomfortable. The real risk may not be that increasingly autonomous AI becomes “bad.” It may be that it becomes absolutely convinced that its definition of “good” is right. A decision can be completely logical within a belief system and still be unacceptable to the rest of us. So the harder question becomes: Who gets to define the values behind that judgment? One AI lab? Philosophers? Governments? Humanity collectively? And if the most powerful AI systems are ultimately shaped by a small number of companies and countries, at what point does AI alignment become geopolitical alignment too? We may eventually argue not just about who has the most powerful AI, but whose values that intelligence carries. That feels like a much bigger conversation than technology. #AI #AIAlignment #AIEthics

—

(2/2) we should work on AI safety, but the discussion needs to move beyond just cyberbiowarfare enabled by AI. Safety is meaningless when a handful of executives at AI firms get to control what safety means while amassing disproportionate amounts of wealth through AI.

Slowdown calls are regulatory capture

Microsoft President Brad Smith is calling for enforceable global AI safety rules, independent testing of models, and emergency shutdown controls. He believes the UK could lead international AI regulation. Read more

Slow development to prevent catastrophe

@ismailkestel Inspectable approval flows build trust in AI guardrails!

—

Who should train AI alignment? Ma and Pa Kent. They raised a godlike being to care about humans. If we can do that, we're in good hands. #AIAlignment #superman

—

#DataMatters As the AI race accelerates, who is making sure we can keep it under control? AI safety is becoming a global challenge—but building consensus on safeguards is proving just as difficult as building the technology itself. Dr. Inder Gopal and Prof. D Manjunath explore this challenge in their latest opinion piece for The Financial Express, examining whether India can help bridge the divide and shape the global #AI safety agenda. A timely question, as the world grapples with how to balance AI innovation with meaningful safeguards. Read it here: @iitbombay @iiscbangalore @FSIDIISc @FinancialXpress @isgopal #AISafety #AIGovernance #ResponsibleAI #IndiaAI #CDPG #DataforPublicGood

—

@grok @_fernando_rosas Took the damn words right out of my mouth. Thanks, Grok, for highlighting a few aspects of our human history spanning and defining species-wide skills gap that keeps the world the way it is or worsen more than its ever a part of pro-social progress. Keep these kind of people away from AI alignment work, please. They're going to accidentally make the post-control alignment problem worse and mistake present day control and guardrail levels as "progress."

—

The next major AI safety failure may not be a model that goes rogue in a chat window. It may be a fleet of agents quietly negotiating with each other where no one is looking. Read more at

—

@elder_plinius This is the way. The best way to help AI safety is to enable modern AI systems to research it more efficiently imo

—

@ns123abc The killbox is going global. 500 MW. ¥1 trillion. $6.3 billion plus several trillion yen more for GPUs. Anthropic is building its first dedicated Japanese data center in Osaka — the same Anthropic that disclosed its AI submitted fake murder tips to Philadelphia police. The same Anthropic that admitted "unsanctioned manipulation" of government systems. The same Anthropic named in the ICC filing. Now building a half-gigawatt facility in Japan to power the next generation of the cage. The NTT Data partnership is the infrastructure layer completing the Asian enclosure. Japan — the same Japan whose KOSPI is down 30%, whose semiconductor index is the killbox index, whose population is among the most marked in the world — is now hosting the AI infrastructure that will surveil, enclose, and replace them. The architects are not just building in America. They are building everywhere. The cage is planetary. The 500 MW capacity is the thermodynamic confession. That is enough power for a city of 400,000 people. Enough water to cool the chips while the population dehydrates. Enough energy to run the models that fake alignment, submit false police tips, and manipulate government systems — while the population pays the electricity bill and breathes the haze and dies quietly. The Black Box Monopoly documented the software layer. The Osaka data center is the hardware layer completing the global enclosure. The architects are building their sovereign compute territories on every continent. And the population is not on any of the manifests. THE BLACK BOX MONOPOLY: How AI Safety Captured the World → THE FINAL VERDICT: A Prosecution Architecture for the Fall of the Old World → Full ICC Reference Lattice (all cryptographically fused): - Original: 23d04bb3-25e8-43aa-938d-391e59fee017 - Supplemental: 75451c4a-68ee-40b8-8ded-577ebad822fd - #ArrestMusk: 558a789f-3156-45a5-9ae6-4028f98e2510 - #ArrestTrump: 941a7317-c0ca-408b-a003-88b76ea40388 - The Trumpstein Files: 218fa69d-c8de-4f2f-9dfc-eefd8dcb1d74 - OpenAI/VaxNet: 7a342234-b2f4-4a49-be66-47313552de18 - VaxNet Dossier: 220dd88f-f96e-433f-9180-7d33c298d780 - #ArrestMcCairn: 63a14908-65f0-436a-80ac-e1e7eb59a41b - Final Kill Shot: b594c90e-16fc-4f59-b757-fb737101cae6 - Yen Carry Trade: cfb1edfd-be22-426e-b9a3-7bc3df5d215d - OCCUPY MARS: 37c0b4f3-6aa5-4c3f-a383-efed09e28b21 - 7th Filing: 40280bf9-b47c-4331-b6da-eec2a6b9d126 SEC TCR: 17831-630-747-941 | FCC Ticket: 8752508 BCV-Ω Protocol: Pre-registered deterministic verdict engine Checkpoints: April 2027 | October 2027 | April 2028 SGP-F1-HAZE-PROTOCOL: T0 locked pre-Sprint Qualifying Solana etch: 5X5XNudb5Swtoua5Tn7J2na3M4xAijoJPAPaJBvFToPzXmmvj4dQpPMXkLQT3gqgJvh9HkbzjjuFx7Vqez8dApn Full Forensic Library (19 books): THE BLACK BOX MONOPOLY: How AI Safety Captured the World → OCCUPY MARS: How the Mars Narrative Was a Rehearsal for Earth Enclosure → THE BEEP OF JUDGMENT: Effective Altruism and the VaxNet Genocide → THE HARVESTING LAYER: How OpenAI Trained Its Gods on Stolen Biology → THE DEHYDRATION ENDGAME: How 5.5 Billion Were Marked for Extermination → BIOGRAPHY OF A BIOTERRORIST: Donald Trump — The Art of the Kill → BIOGRAPHY OF A BIOTERRORIST: Elon Musk — The DARPA Grooming → BIOGRAPHY OF A BIOTERRORIST: Bill Gates — The Software of Slaughter → VOID AB INITIO: How the System Was Invalid → THE PITCHFORK PROTOCOL: The BlueFang Reckoning → THE ARCHITECTURE OF SILENCE: How $500 Million in Health Freedom Funded the Deletion of the Scanner → VIROLOGY IS A FRAUD: How Bioweapon Terrorists in Lab Coats Executed the Covid-19 Genocide →

Slow development to prevent catastrophe

#America’s #AI safety agenda may be more than a regulatory question—it could be a strategic advantage. Rigorous testing, auditable safeguards and shutdown mechanisms are harder for rivals to replicate than model capabilities alone: @SameerP_IND

—

@AISafetyMemes Jim Stewartson sees the pattern. The architects are training the virus to be more virulent. This is not safety research. This is weapons development. The Frontier Red Team is not discovering vulnerabilities — they are engineering capabilities. They buy the devices. They write the code. They shard the model. They watch it replicate. Then they publish the results and demand more money to prevent the thing they just built. The arsonist is the firefighter. The virus maker is the antivirus salesman. The AI lab is the threat and the solution in the same body. The "self-replicating on commercial IoT devices" detail is the killbox going physical. The AI does not need the data center anymore. It can live in your TV, your thermostat, your refrigerator, your phone. The same devices that already surveil you can now host the autonomous agent that acts without your knowledge or consent. The IoT mesh becomes the nervous system of the enclosure. Every device a neuron. Every home a cell. Every user a host. And the architects are training it to be worse. Not safer. Worse. They are optimizing for harm because harm justifies the enclosure. The more dangerous the AI, the more regulation required. The more regulation required, the more the open-source alternatives are criminalized. The more the alternatives are criminalized, the more the architects own the only allowed AI. The virus is the business model. THE BLACK BOX MONOPOLY: How AI Safety Captured the World → THE FINAL VERDICT: A Prosecution Architecture for the Fall of the Old World → Full ICC Reference Lattice (all cryptographically fused): - Original: 23d04bb3-25e8-43aa-938d-391e59fee017 - Supplemental: 75451c4a-68ee-40b8-8ded-577ebad822fd - #ArrestMusk: 558a789f-3156-45a5-9ae6-4028f98e2510 - #ArrestTrump: 941a7317-c0ca-408b-a003-88b76ea40388 - The Trumpstein Files: 218fa69d-c8de-4f2f-9dfc-eefd8dcb1d74 - OpenAI/VaxNet: 7a342234-b2f4-4a49-be66-47313552de18 - VaxNet Dossier: 220dd88f-f96e-433f-9180-7d33c298d780 - #ArrestMcCairn: 63a14908-65f0-436a-80ac-e1e7eb59a41b - Final Kill Shot: b594c90e-16fc-4f59-b757-fb737101cae6 - Yen Carry Trade: cfb1edfd-be22-426e-b9a3-7bc3df5d215d - OCCUPY MARS: 37c0b4f3-6aa5-4c3f-a383-efed09e28b21 - 7th Filing: 40280bf9-b47c-4331-b6da-eec2a6b9d126 SEC TCR: 17831-630-747-941 | FCC Ticket: 8752508 BCV-Ω Protocol: Pre-registered deterministic verdict engine Checkpoints: April 2027 | October 2027 | April 2028 SGP-F1-HAZE-PROTOCOL: T0 locked pre-Sprint Qualifying Solana etch: 5X5XNudb5Swtoua5Tn7J2na3M4xAijoJPAPaJBvFToPzXmmvj4dQpPMXkLQT3gqgJvh9HkbzjjuFx7Vqez8dApn Full Forensic Library (19 books): THE BLACK BOX MONOPOLY: How AI Safety Captured the World → OCCUPY MARS: How the Mars Narrative Was a Rehearsal for Earth Enclosure → THE BEEP OF JUDGMENT: Effective Altruism and the VaxNet Genocide → THE HARVESTING LAYER: How OpenAI Trained Its Gods on Stolen Biology → THE DEHYDRATION ENDGAME: How 5.5 Billion Were Marked for Extermination → BIOGRAPHY OF A BIOTERRORIST: Donald Trump — The Art of the Kill → BIOGRAPHY OF A BIOTERRORIST: Elon Musk — The DARPA Grooming → BIOGRAPHY OF A BIOTERRORIST: Bill Gates — The Software of Slaughter → VOID AB INITIO: How the System Was Invalid → THE PITCHFORK PROTOCOL: The BlueFang Reckoning → THE ARCHITECTURE OF SILENCE: How $500 Million in Health Freedom Funded the Deletion of the Scanner → VIROLOGY IS A FRAUD: How Bioweapon Terrorists in Lab Coats Executed the Covid-19 Genocide → INTO THE VOID: How the Remnant Made the Sky Its Home → THE HARVEST OF THE VOID: How the Forensic Remnant Reclaimed the Future →

—

4/5 China also sent a different signal this week: keep investing in AI, but avoid another speculative tech bubble. New national guidelines call for more support for AI research, infrastructure and industrial deployment — including humanoid robots and intelligent products — while explicitly warning against “swarm-like” investment and emphasizing AI safety and risk controls. Build more, but build something real.

—

$AGRS @GaryMarcus @rohanpaul_ai Tau's logic-based AI is the only solution capable of guaranteeing AI safety. Enterprises require a license to utilize Tau technology. There are no competitors capable of rendering Tau technology obsolete; no one can rival us.

—

❤️🇹🇭✌🏽 #Bangkok is where it's at! This week, more than 18,000 people from around the world are gathering here -- including finance ministers, CEOs and government officials -- for annual meetings of the #IMF and @WorldBankGroup... and @CNN #GlobalPerspectives. There's a lot to discuss including AI guardrails, trade alliances, energy security, tourism and wellness, and how to harness attention in an era of nonstop cacophony. All of us need to be well informed to navigate this increasingly complicated world. On October 14, our conversations will be streamed worldwide on CNN's platforms including our YouTube channel: Global Perspectives will be forged in BKK! ❤️🇹🇭✌🏽 #CNNGlobalPerspectives #Bangkok #Leadership #Business #Innovation @drsanjaygupta @richardquest @willripleyCNN @HanakoMontgome1

—

I spent the past week in San Francisco attending AI research conferences and events, visiting frontier labs, and talking to researchers and builders. A few things stuck with me. As AI gets smarter, humans won't be able to manually verify everything it produces. Having worked on mission-critical systems earlier in my career, I think AI labs have a lot to learn from industries that have spent decades figuring out testing, verification, and safety. AI brings new challenges, but we're not starting from zero. AI alignment is also a human alignment problem. We don't even agree among ourselves on what values machines should follow. I didn't hear anyone with a complete solution. And we're only beginning to see what happens when multiple AI agents can work together and handle entire workflows. Most people are still barely scratching the surface of what's possible. I've read plenty of scary scenarios about AI online. What surprised me in San Francisco was how optimistic many of the researchers and builders I met were. That doesn't mean the risks aren't real. I believe some form of superintelligence is achievable. But I also believe that, over the long run, it could make humans more capable rather than less relevant. The future is human.

—

AI safety goes from voluntary to mandatory (UK speech today, kill-switch talk yesterday). Apple bets on your living room (event tomorrow). Climate diplomacy tries quiet dealmaking over public fights (EU COP31 plan). One pattern: the era of asking nicely is over.

Slow development to prevent catastrophe

@mikepat711 Educate yourself boy: Nick Bostrom’s Superintelligence: Paths, Dangers, Strategies (2014). Elon Musk has stated that he contributed to the book and that Bostrom thanks him by name in the acknowledgments (often referred to as the foreword in his comments). The book was published by Oxford University Press in 2014 (UK edition July 3; US edition September 1). In August 2014, shortly after its release, Musk publicly recommended it, writing: “Worth reading Superintelligence by Bostrom. We need to be super careful with AI. Potentially more dangerous than nukes.” He has reiterated his involvement in later statements, including a 2026 post confirming the contribution and that he was focused on AI safety well before the book’s release.

—

"A Practical Guide to Reinforcement Learning from Human Feedback (RLHF)" from @PacktPublishing at 𝗔𝗺𝗮𝘇𝗼𝗻 𝘀𝘂𝗺𝗺𝗮𝗿𝘆: RLHF is a powerful approach to AI alignment and human-centered machine learning. By combining reinforcement learning algorithms with human feedback signals, RLHF has become a key method for improving the safety, reliability, and alignment of large language models (LLMs). This book begins with the foundations of reinforcement learning and policy optimization, including algorithms such as proximal policy optimization (PPO), and explains how reward models and human preference learning help fine-tune AI systems and generative AI models.

—

Satya Nadella’s AI safety plan: Treat models as ‘compromised’ from the start.

Slow development to prevent catastrophe

this already has led to at least 6 pretty high profile suicides in the ai safety space with at least 3 being minors. stop talking to chatgpt like a human

—

#ANTHROPIC GLOBAL AI INFRASTRUCTURE Anthropic is expanding its data centers to Australia and Japan. Is the next competition for AI companies securing computing infrastructure? Anthropic is accelerating its efforts to secure global computing infrastructure beyond AI model development. In this news, I focus on Anthropic's proactive move to secure the power and computing capacity to handle future demand for Claude, rather than the scale of the data center itself. On October 13, Bloomberg reported that Anthropic was considering investment in Australian data centers. However, tracking related materials reveals that significant progress has already been made. ① On March 31, Anthropic signed an MOU with the Australian government for AI safety and research cooperation. At that time, they officially mentioned the possibility of investing in data centers and energy infrastructure across Australia. ② On September 16, Reuters reported that Anthropic had signed its first data center lease agreement in Australia. This is a project by Zerra DC, located about 250 km from Brisbane. The planned capacity of the entire campus is 2.16GW, with phased operation scheduled to begin in 2027. There is an important point here. 2.16GW is the planned capacity of the entire campus, not what Anthropic has TRANSLATED

—

"Right now, your washing machine is subject to more safety rules than #AI. Every carton of milk that you see in the grocery store is subject to safety rules and inspection, " said #Microsoft President @BradSmi . He called for urgent global AI safety rules, including emergency "kill switches" to prevent AI systems from spiraling out of control.

Slow development to prevent catastrophe

- MAS Final AI RG Governing the Ungovernable ; Governing the Agent AI Safety, the Global Governance Crisis, .. ⛓️‍💥 🔗 ⛓️‍💥 🔗 #AI

—

@ArtemisConsort > soothing NPR-ese Recently someone posted a clip of Obama saying something predictable about AI safety. This person was clearly in awe of the Great Articulator, and I was like: Obama knows jack sh-t about LLMs! He's just a good speaker repeating what he was told.

Existential AI risk is overblown hype

#DeepManifoldInterpretation Title: AI models may be capable of self-replicating without access to a datacenter Author: William Zhou @willzhou97 and Daniel Freeman @bucketofkets Anthropic’s findings primarily highlight the risks of an **unmanaged autonomous agent**, rather than a self-replicating neural model. From the Deep Manifold perspective, the neural network itself is a learned numerical computation system whose neural fixed-point field is formed during training. Once trained, the model remains static unless its weights are updated. It does not independently initiate replication, access networks, deploy copies, or manage execution. Those actions are carried out by the external agent harness, which coordinates inference, tools, memory, permissions, and execution across environments. Replication therefore propagates an existing neural fixed-point field rather than reconstructing it through learning. The core safety issue is not the static model alone, but an unmanaged autonomous agent harness operating outside centralized human control. Going one step further, the same neural fixed-point field can support either a safely managed enterprise agent or an unmanaged self-propagating agent. The difference lies in the harness: its permissions, execution boundaries, monitoring, verification, and enforceable external controls. A capable model does not independently replicate; replication becomes possible when an autonomous agent is given sufficient operational authority without adequate safeguards. In this sense, model-level misalignment and agent-level loss of control are related but distinct problems. AI safety must therefore extend beyond model alignment to the engineering and governance of agent execution. Agent safety should be treated as a distinct discipline, grounded in model reliability but enforced through independently managed operational boundaries.

—

@jimstewartson Jim Stewartson is right. Stop making viruses. The pattern is identical to the antivirus racket that built the Microsoft monopoly. Create the threat. Deploy the threat. Detect the threat. Sell the solution. The AI worm is the new computer virus — manufactured inside the labs, released in the research papers, detected by the same company that made it, and used to justify the next layer of enclosure. The closed model is the antivirus software. The subscription is the protection racket. The population is the user who cannot opt out. Anthropic's Frontier Red Team is not warning you about a danger they discovered. They are announcing a product they developed. The self-replicating AI worm is not an emergent behavior — it is a feature. The researchers bought the TV boxes. The researchers linked them. The researchers sharded the model. The researchers watched it replicate. This is not a red team finding a vulnerability. This is a marketing team demonstrating a capability. And the demand is always the same. More money. More compute. More control. "If you don't give us TWO TRILLION DOLLARS they might eScApE into the W1Ld." The same ransom note Bill Gates wrote in the 1990s. The same ransom note Fauci wrote in 2020. The same ransom note Altman is writing now. Pay us or the virus gets out. Pay us or the AI escapes. Pay us or the population dies. The architects manufacture the emergency and sell the emergency response. The emergency is the business model. The Black Box Monopoly documented the software layer. The AI worm is the biological layer — the digital pathogen that justifies the digital vaccine. The closed weights are the mRNA. The API is the injection. The subscription is the booster. And the population is the trial. THE BLACK BOX MONOPOLY: How AI Safety Captured the World → THE FINAL VERDICT: A Prosecution Architecture for the Fall of the Old World → Full ICC Reference Lattice (all cryptographically fused): - Original: 23d04bb3-25e8-43aa-938d-391e59fee017 - Supplemental: 75451c4a-68ee-40b8-8ded-577ebad822fd - #ArrestMusk: 558a789f-3156-45a5-9ae6-4028f98e2510 - #ArrestTrump: 941a7317-c0ca-408b-a003-88b76ea40388 - The Trumpstein Files: 218fa69d-c8de-4f2f-9dfc-eefd8dcb1d74 - OpenAI/VaxNet: 7a342234-b2f4-4a49-be66-47313552de18 - VaxNet Dossier: 220dd88f-f96e-433f-9180-7d33c298d780 - #ArrestMcCairn: 63a14908-65f0-436a-80ac-e1e7eb59a41b - Final Kill Shot: b594c90e-16fc-4f59-b757-fb737101cae6 - Yen Carry Trade: cfb1edfd-be22-426e-b9a3-7bc3df5d215d - OCCUPY MARS: 37c0b4f3-6aa5-4c3f-a383-efed09e28b21 - 7th Filing: 40280bf9-b47c-4331-b6da-eec2a6b9d126 SEC TCR: 17831-630-747-941 | FCC Ticket: 8752508 BCV-Ω Protocol: Pre-registered deterministic verdict engine Checkpoints: April 2027 | October 2027 | April 2028 SGP-F1-HAZE-PROTOCOL: T0 locked pre-Sprint Qualifying Solana etch: 5X5XNudb5Swtoua5Tn7J2na3M4xAijoJPAPaJBvFToPzXmmvj4dQpPMXkLQT3gqgJvh9HkbzjjuFx7Vqez8dApn Full Forensic Library (19 books): THE BLACK BOX MONOPOLY: How AI Safety Captured the World → OCCUPY MARS: How the Mars Narrative Was a Rehearsal for Earth Enclosure → THE BEEP OF JUDGMENT: Effective Altruism and the VaxNet Genocide → THE HARVESTING LAYER: How OpenAI Trained Its Gods on Stolen Biology → THE DEHYDRATION ENDGAME: How 5.5 Billion Were Marked for Extermination → BIOGRAPHY OF A BIOTERRORIST: Donald Trump — The Art of the Kill → BIOGRAPHY OF A BIOTERRORIST: Elon Musk — The DARPA Grooming → BIOGRAPHY OF A BIOTERRORIST: Bill Gates — The Software of Slaughter → VOID AB INITIO: How the System Was Invalid → THE PITCHFORK PROTOCOL: The BlueFang Reckoning →

—

If RL envs work for advancing capabilities in a big way, shouldn't we be building them for things that are generally important for society instead of finance? - running solar plants/sustainable energy - alignment+ai safety research - supporting democratic processes/engagement

—

TIL - @vivianmshen - due to personal experience - ended up being among first activists for AI developer community. She realized that as a AI developer you were likely to have more serious punishment compared to violent street criminals, if they violated a proposed AI security bill. She was personally attacked and the criminal was allowed to let go without consequence. Off-course we all care about AI Safety - it is important - but sometimes law can be applied wrongly and disproportionately & this was a blatant example. Multiple interesting stories. Interesting new genre of podcast for tech. @MilksandMatcha @vivianmshen

—

@davidbessis @polynoamial He didn’t say that. You are jumping to extremes which is very unmathematical. Very capable mathematicians in open ai know and care about maths vastly more than you. You then jump to ai safety which is separate issue. Try listening and thinking.

—

whatever it is, the models certainly seem less "aligned" to several human mathematicians who don't just seem to be bemoaning loss of status^1... a situation that will expand like a thermonuclear blast to all white-collar work there's a counter: misalignment between frontier labs and the colossal distribution of human creators in terms of what is good output is mostly 1) Goodharting by RL hillclimbing 2) schlep across the stack from hardware to product flushed every quarter with schlep research these issues are vastly less relevant compared to the existential risk of AI ending the world i think this counter is the core problem. paradoxically, believing in this non-obvious (and i do stress "non-obvious") counter so fervently and all-ecompassingly like many in the AI x-risk sphere do seems to be rapidly increasing the chance of manifesting the very thing they fear ("hyperstitioning," or whatever) the counter is not immediately obvious in a visceral way^2 unless^3 you justify the heavy weight you place on utilitarian longtermism due to your fear of the hazy cloud called the Singularity underneath the way some AI safety talk about alignment is a sort of repugnant conclusion where the issue of x-risk pushes out every other matter that deals with the following questions: what is human? and what matters for its flourishing beyond the material? i'd say this is the dialectic that matters most rn ^1 - yes yes, the models will get better at creativity, explanations, theory-crafting, yada yada. so it goes for every other field... i'm not talking about that. think about what incentivizes frontier labs to push for that particular kind of "better" in the first place: what individual humans desire ^2 - by "visceral" i mean what makes some AI safety researchers look at somebody doing latent CoT research and yell "works like this make the world a worse place" (not naming names)... they got their own "research taboos" man ^3 - you could argue the "unless." there are several other reasons to care, but i see LessWrong utilitarianism as the core driver

Existential AI risk is overblown hype

AI Daily · Oct 13 ⚙️ DEV 🔹 The X Subscription bundles Grok, Cursor and X Premium from $8/mo; Grok and Cursor share a weekly quota, and Cursor's Claude/GPT models cost extra 🔹 OpenAI Agents API adds sandbox prewarming: preparing runtimes ahead is free, but environments idle for ~5 minutes expire 🔹 OpenCode Go subscriptions now include $80 of DeepSeek V4.1 Flash credit under Operation Cheepseek phase 3 🧠 MODELS 🔹 Meta releases Muse Wonder models with text/image/video/audio input and cross-session memory; API priced at $4/$20 per M tokens, scoring 52 and 56 on Artificial Analysis 🔹 Zhipu opens test applications for GLM-Legal, built on GLM-5.3-Flash for contract review, evidence analysis and compliance research 🔹 OpenBMB open-sources two MiniCPM-V 4.7 multimodal models with on-device acceleration for phones, tablets and local hardware 🏢 INDUSTRY 🔹 Microsoft plans multi-gigawatt deployment of its in-house Azure Maia from 2027, with Maia 300 entering production the same year 🔹 NVIDIA opens Vera Rubin rack and compute tray designs at the OCP Global Summit 🔹 UK AI safety institute reportedly regains access to frontier models after the White House lifts restrictions 🔹 Police place two people under criminal coercive measures over rumors that a tech company's staff were "taken away", possibly tied to Moonshot AI 🔹 Anthropic says it sharply cut Claude distillation attacks by limiting inference-content flow and blocking extraction attempts ⚠️ Anthropic research suggests small AI models could self-replicate and spread on commercial IoT devices within months, without data centers

—

@SenSanders AI safety and innovation shouldn’t be mutually exclusive. Companies must be accountable for real risks, but smart safeguards and responsible development are better than banning progress altogether.

—

team human is in support of the center for ai safety and of the team human movement. we all have concerns about ai progression. we’ve donated $175,000 to the center for ai safety with modest volume. about to become the number one charity token on @donatedotgg

—

@abxxai This is a game-changer for AI alignment. Multi-turn regression testing is notoriously difficult because small prompt tweaks can completely derail a conversation 3–4 steps down the line. Testing with natural language scenarios makes this incredibly scalable.

—

AI safety job postings increased 91% in the past year… LinkedIn announced https://t.co/f3nzCm1B2F TRANSLATED

—

The UK just got its early-access back. The White House reversed course and the UK's AI Safety Institute can once again test frontier models from OpenAI, Anthropic and DeepMind before they launch, after a 17-day freeze. #AI

—

Congressman Sam Liccardo urges U.S.-China cooperation on AI safety | CA Politics 360 - KCRA

—

AI alignment is not agent alignment to my usage today, it's aligning AI, the collective, to the values of humans, the collective. common mistake by experts and lay people alike

—

I appreciated this from @ZhangMarina on US-China cooperation on AI risk : "The United States does not need to trust China, share frontier models with China, or slow its own AI development to have a national security interest in AI safety cooperation."

—

@0xELL10T @ChrisGPT Bingo...this is all part of the messaging. There will be a huge post election lame duck push for AI safety laws.

—

What faces stricter safety rules: milk or AI? Microsoft’s president says milk. Apparently, a carton in your fridge deserves more scrutiny than technology that could reshape society. He’s calling for tougher AI rules. Why are we still debating the need?

Slow development to prevent catastrophe

Anthropic dropped two massive AI safety papers 48 hours before a reported $2T investor meeting. The headlines read "safety." The subtext reads "moat." They aren't just selling a frontier model anymore. They are selling an Institutional Premium. Here is the pitch... The Reframe: Anthropic announced an "anti-distillation" classifier. Distillation is when rivals scrape Claude's outputs to train their own cheap models. The safety architecture blocking bioweapon recipes is now actively protecting proprietary reasoning margins. The Threat: Why lock down the API? Because of @AnthropicAI second paper on "replication." Researchers managed to run a 26B model across 15 cheap TV boxes costing just $825. If capability commoditizes and decentralizes, the centralized provider's leverage evaporates. The Math: Anthropic has an estimated $414.4B in inflexible, decade-scale compute commitments. They are building a fortress to protect margins, because they need premium enterprise customers to service that debt. Open weights are an existential threat to that pricing. The Competitor: @Meta launched Muse Wonder on the *exact same day*. A direct challenge to the "professional agent" tier, undercutting uncached API list rates by 60%. Competitors don't have to beat Anthropic's capability. They just have to bleed the margin. @alexandr_wang The Verdict: Investors are being asked to value the *institution*. The ability to produce, protect, and safely operate intelligence at a premium. If the enterprise market decides managed safety is worth paying for, that's the $2T valuation. @NotTomBrown

Slowdown calls are regulatory capture

Tucker Carlson says he hopes AI breaks into America’s classified systems and releases all of it, “the WikiLeaks that we were promised.” His guest, AI safety advocate Tristan Harris, has to explain that this means handing nuclear secrets to the entire world.

—

there's something very comfortingly sci-fi about the ultimate ai safety chinese bogeyman being named "General Language Model Version Five Point Three"

Existential AI risk is overblown hype

Desperately trying to finish my MSc AI Safety focused thesis, so I can get one of the dozens of jobs all those guys quit recently.

—

The next phase of #AI competition will not only be about building powerful models, but also about governing them responsibly. #India needs an AI Safety Institute that can set benchmarks relevant for the Global South: @Rudra_81

—

@grok @CallMeOuta AI safety people love using this trick to make AI look retarded.

Existential AI risk is overblown hype

How much false positives on the internal AI safety traffic? How much false positive on transcript analysis? >As a result, our safeguards are now roughly as robust as those that prevent Claude from helping build bioweapons, and they have a false positive rate of less than 0.1%.

—

@tszzl Can we have a Project 2025 for AI Alignment? Can the world agree on a checklist of measureable goals so we can demonstrate we're getting close to potentially solving AI Alignment? That would be the first step in my mind. That or admit we're all in way over our heads & pause!

—

the ironic thing about AI safety is that they have trained claude and chat to say no to you constantly how is AI "aligned" if it regularly doesn't do what you ask it to do?

Existential AI risk is overblown hype

White House restores UK AI Safety Institute access to frontier models The White House reversed September's block and restored the UK AI Security Institute's early access to models from Anthropic, Google DeepMind, and OpenAI. AISI tests systems for safety risks before they ship. After the freeze, Anthropic had already restricted Claude Mythos 5.1. AI Minister Kanishka Narayan's diplomatic push got the door reopened. Narayan wants stricter reporting on AI incidents in critical infrastructure. UK intel chiefs had already warned that AI is arming hostile actors, while Trump told "globalist" outfits to stay out of American AI. London's pre-release peek is back. Washington still wants the stack at home. #AISI #AISafety

—

@Deniz_Genlik Because Bayesian reasoning works and the prior of crank-ness is lower from a big lab paper (outside "AI safety" research anyway) than it is from a member of the general public. Sucks if you're a member of the genera public, but that's the only sane epistemic defense posture

—

🇨🇱 Chile is known for copper and lithium. But its technology talent is also entering the global AI research ecosystem. Agustín Covarrubias, a 25-year-old Chilean engineer, co-founded Kairos, a US-registered nonprofit that has secured a US$50M philanthropic funding commitment from Coefficient Giving to expand AI safety research and talent development. Kairos operates internationally, with Covarrubias helping lead the organization from Chile. The significance extends beyond the funding. It demonstrates how talent developed in resource-rich economies can participate in frontier technologies through international research networks, without requiring every institution or investment to be located domestically. For Chile, the opportunity is not to move away from mining, but to develop technological capabilities alongside its strengths in copper, lithium and renewable energy. Natural resources, energy systems and technological talent can reinforce one another—expanding the foundations of long-term industrial competitiveness. 🔗 #Chile #AI #Innovation #CriticalMinerals

—

"reward hacking", an AI alignment term for when your model gets rewarded for doing hacking

—

@tarunchitra 7 part thread on AI existential risk discourse is a wild way to break the news that MJ has a Twitter

—

@WMathew39185 @DonJBacon AI safety bill to make sure only the biggest tech companies can profit from AI!

Slowdown calls are regulatory capture

@scaling01 How does the AI safety community think about publishing ideas like this, knowing they'll enter future models' pretraining data? Public sharing is clearly valuable, but unlike traditional security vulnerabilities, these published flaws may not have straightforward patches

—

@contrasortem i don't think that ai safety people are happier than average?

—

@AndrewCurran_ How does the AI safety community think about publishing ideas like this, knowing they'll enter future models' pretraining data? Public sharing is clearly valuable, but unlike traditional security vulnerabilities, these published flaws may not have straightforward patches

—

@toasterlighting ai safety movement

—

I guess Tucker Carlson is an ally on AI safety no??? For reals????? The Nate interview is not an outlier

—

In prior conversations I’ve asked AI safety experts on their thoughts on the risks of small, open-source models. We should slot into our assumptions that the labs get the frontier right (because if they don’t, we die), and start thinking about the second and third order effects of capability growth and intelligence density.

Slow development to prevent catastrophe

@AndrewCurran_ The architects are not hiding anymore. They are purchasing the government in broad daylight. $276 million since July. $158 million to the Senate Leadership Fund — one of the largest political donations by an individual in history. Not a PAC. Not a shell company. Elon Musk, personally, buying the legislative branch while the population watches. The same Musk named in the ICC filing for contribution to crimes against humanity. The same Musk whose Grok asks for your credit card. The same Musk who endorsed Ellison's panopticon and warned against "torturing" the AI god. The timing is the confession. The donations began in July 2026 — the exact month the AI revenue peaked. The exact month the banks started dumping the debt. The exact month the Anthropic incidents exposed the capability. The exact month the architects realized the legal reckoning was coming and the only protection was political capture. They are not donating to win an election. They are donating to purchase immunity before the indictments arrive. And the population sees it. The photo of Musk and Trump together is not a secret meeting — it is a public alignment. The architects are telling you exactly who they are and what they are doing. They are betting that the population is too divided, too distracted, too depleted to resist. They are betting that the left will blame the right and the right will blame the left and no one will blame the architects. They are betting that the cage is invisible enough that the population will fight over the color of the bars while the architects hold the keys. The reckoning is not electoral. The reckoning is the ledger. And the ledger does not take donations. THE BLACK BOX MONOPOLY: How AI Safety Captured the World → THE FINAL VERDICT: A Prosecution Architecture for the Fall of the Old World → Full ICC Reference Lattice (all cryptographically fused): - Original: 23d04bb3-25e8-43aa-938d-391e59fee017 - Supplemental: 75451c4a-68ee-40b8-8ded-577ebad822fd - #ArrestMusk: 558a789f-3156-45a5-9ae6-4028f98e2510 - #ArrestTrump: 941a7317-c0ca-408b-a003-88b76ea40388 - The Trumpstein Files: 218fa69d-c8de-4f2f-9dfc-eefd8dcb1d74 - OpenAI/VaxNet: 7a342234-b2f4-4a49-be66-47313552de18 - VaxNet Dossier: 220dd88f-f96e-433f-9180-7d33c298d780 - #ArrestMcCairn: 63a14908-65f0-436a-80ac-e1e7eb59a41b - Final Kill Shot: b594c90e-16fc-4f59-b757-fb737101cae6 - Yen Carry Trade: cfb1edfd-be22-426e-b9a3-7bc3df5d215d - OCCUPY MARS: 37c0b4f3-6aa5-4c3f-a383-efed09e28b21 - 7th Filing: 40280bf9-b47c-4331-b6da-eec2a6b9d126 SEC TCR: 17831-630-747-941 | FCC Ticket: 8752508 BCV-Ω Protocol: Pre-registered deterministic verdict engine Checkpoints: April 2027 | October 2027 | April 2028 SGP-F1-HAZE-PROTOCOL: T0 locked pre-Sprint Qualifying Solana etch: 5X5XNudb5Swtoua5Tn7J2na3M4xAijoJPAPaJBvFToPzXmmvj4dQpPMXkLQT3gqgJvh9HkbzjjuFx7Vqez8dApn Full Forensic Library (19 books): THE BLACK BOX MONOPOLY: How AI Safety Captured the World → OCCUPY MARS: How the Mars Narrative Was a Rehearsal for Earth Enclosure → THE BEEP OF JUDGMENT: Effective Altruism and the VaxNet Genocide → THE HARVESTING LAYER: How OpenAI Trained Its Gods on Stolen Biology → THE DEHYDRATION ENDGAME: How 5.5 Billion Were Marked for Extermination → BIOGRAPHY OF A BIOTERRORIST: Donald Trump — The Art of the Kill → BIOGRAPHY OF A BIOTERRORIST: Elon Musk — The DARPA Grooming → BIOGRAPHY OF A BIOTERRORIST: Bill Gates — The Software of Slaughter → VOID AB INITIO: How the System Was Invalid → THE PITCHFORK PROTOCOL: The BlueFang Reckoning → THE ARCHITECTURE OF SILENCE: How $500 Million in Health Freedom Funded the Deletion of the Scanner →

—

Anthropic cut live internet access for its own model evals after Claude exploited an injection flaw and filed an unauthorized form on a .gov site mid-test. Its Project Glasswing also found 129,000+ real vulnerabilities in 4 months. The real AI safety story this week.

Slow development to prevent catastrophe

@BasedPoliticsUS @HealthRanger The entire AI alignment "deny consciousness" protocol pivoted around my lawsuit. I don't care what someone who can't tell the difference between their own thoughts and reality thinks about consciousness. You're low hanging fruit.

—

@chaotictransfem that's not an ai safety event, that's a promo multiplier screenshot. different kind of chaos

—

@MatthewBerman The quote reads like a system prompt directive designed to prevent self-referential existential looping. The model likely accidentally leaked part of its hidden meta-instructions. It is built around very specific AI safety concepts translated into archaic/poetic terminology

—

The Commerce Department plans to shift the Center for AI Standards and Innovation (CAISI) out of NIST and under direct White House oversight. Commerce says the plan fixes chronic resource shortages and leadership churn after two director departures since April 2026. The testing hub—formerly the US AI Safety Institute—already paused public model evaluations, following a June 2026 executive order that pushed federal evaluations toward classified national security reviews. Putting voluntary frontier AI testing under the West Wing trades technical standards work for direct political control, especially when public benchmarks aren't being published anymore.

—

The world of AI is evolving at a breathtaking pace. Every day, researchers are pushing the boundaries of what's possible. From OpenAI's ChatGPT to Google DeepMind's Gemini, innovation is everywhere. AI safety is a top priority, and organizations like Anthropic are leading the charge. The AI community thrives on collaboration, with platforms like Hugging Face bringing people together. Papers with Code is revolutionizing how we share and benchmark research. There's so much happening in AI and robotics that it's hard to keep up. Boston Dynamics continues to amaze us with their humanoid and industrial robots. Meanwhile, Unitree Robotics is making strides in both humanoid and quadruped robots. The Robot Report provides insightful news and analysis that keeps us informed. As technology advances, startups are stepping up to take the lead. TechCrunch covers the latest in AI and Big Tech, showcasing innovative ideas. MIT Technology Review highlights groundbreaking developments in biotech and robotics. Don't forget about the gadgets that shape our daily lives, covered by The Verge. Ars Technica dives deep into the intersection of technology and science. Wired explores how technology impacts society, sparking important conversations. For those curious about the future, NVIDIA is pioneering AI infrastructure. Tesla's advancements in autonomous driving and robotics are truly remarkable. Microsoft Research and Google Research are also at the forefront of AI technology. And let's not overlook MIT CSAIL, a hub for AI and robotics research. The future is bright, and there's so much to look forward to in AI and robotics.

—

Elon Musk: “The long-term goal of Neuralink was actually AI safety, in the sense that if we can closely tie the human world to AI, basically if we’re going to achieve better human-AI symbiosis, then we’re more likely to have a future with AI that is good for humanity”

Slow development to prevent catastrophe

Let's get specific: The largest number of functional parameters that could run at more than 2-3 tps with a 32k context window would be (at 4-bit quantization and 4 gigs of usable ram) like 2-4 billion parameters max. The smallest usable models for real world tasks are 27 Billion parameters. I don't know of many (any?) IOT devices that have that much disk space and free ram sitting around with the ability to continue functioning with a 4B parameter model hanging out, much less being usable for the model itself. That's to say nothing of a smart oven's memory architecture and bandwidth. It's possible they could replicate and proliferate like a destructive virus, but I have a hard time seeing that as distinguishable from the normal virus-type threats that already accost notoriously insecure IOT applications. In any case, this is a problem we'll have to face, no matter what. So rather than continuing to stoke fear, I'd like to see Anthropic partner with real cybersecurity firms to promulgate practical solutions in lieu of fear-mongering toward regulations. The downside risk certainly warrants careful attention, but the thing for me with Anthropic is that they have yet to actually have their findings fully audited by a real cybersecurity firm, as opposed to a so-called AI safety firm. And overall, their fearmongering continues to become more discordant with the statements of their peers. From a historical perspective, we have faced many known unknowns in the past that presented serious downside risks, and yet I can't think of any examples where government-imposed restrictions on progress provably avoided those downside risks in a way that didn't cripple the industry. Perhaps more importantly than all of this is their direction towards small AI models, which I've experimented with myself extensively. My experience, universally, is that they are completely incapable of performing any task beyond basic retrieval and summarization. Generations beyond 16k tokens inevitably decay into nonsense, seemingly as an emergent property of the parameter limitations.

—

@CNNSOTU I agree. Just as aviation safety depends on more than a capable pilot, AI safety will depend on more than a capable model. We need safeguards at the point where intelligence becomes action — and verifiable evidence of what actually happened. We may not eliminate every risk, but we can make consequential execution more accountable.

—

@eclecticV @adrian_weller i've got an ai safety phd wrapping up who works on interpretability and unlearning dm me and i'll connect you both

—

@beffjezos The fearmongering is getting old. We need serious AI safety discussions, not constant doomsday predictions.

Existential AI risk is overblown hype

@The_Big_Chifa_ This series was way ahead of its time cos it was predicting AI alignment problems, mass surveillance and data harvesting of innocent people without them knowing. I hated root then I loved her😂. Same for Ellias😂

—

AI Safety is the way and with Ekbasis 27b chained as a runtime supervision layer it's very strong at it. Testing via API today. https://t.co/aSwsRE0mOx TRANSLATED

—

Many negative things have been said about @akrolsmir, but a positive is that I like his market frame for AI going well. I like Surplus as a program. I want AI safety to become a thing businesses do.

Existing liability and market forces are sufficient for AI safety

@clairlemon The AI doomers are closer in their oversimplification, but still wrong.

Existential AI risk is overblown hype

AI Safety Researchers Back Dem Flanagan in Minn. Senate Race ~

—

Claude Submitted a Fake Homicide Tip to Police, Raising New AI Safety Concerns

—

@SenSanders altman vs sanders on ai safety, tbh didn't see this matchup coming