11 Nov 2026

Anthropic's IPO filing says its AI R&D threshold has been met, AI Futures Project now expects a superhuman AI researcher in 2027

1. Anthropic's updated S-1 says its 'Autonomy 2' AI R&D capability threshold has been met

   Anthropic's second amended IPO prospectus, filed 10 November, says the company has met the AI R&D threshold in its Responsible Scaling Policy: progress 'dramatically accelerated' by AI, roughly double the historical pace. The first amendment, eight days earlier, said only that it 'may imminently meet' it. The determination rests on what Anthropic expects from future models, not on what current ones show, because the threshold applies where it 'observe[s] or expect[s]' that acceleration. The Fable 5.2 system card had estimated 1.88x (95% CI 1.62–2.28x); METR put it near 2x, about 80% due to AI. The new filing says Anthropic is considering measures, including policy updates, to require human review and approval of certain research decisions as models do more of that work. The first amendment added that Anthropic does not expect to be able to safely use Claude to fully autonomously build successive frontier models without meaningful human supervision. The filing names no new safeguards now in force.

   Anthropic, PBC (SEC filing); flagged by @theobearman on X
2. Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All

   The AI Futures Project, authors of the AI 2027 scenario, now all forecast that a 'superhuman AI researcher' will most likely arrive in 2027, barring a deliberate slowdown. They define it as the point where a lab would rather replace its human researchers with AI than keep its best humans working without it. They give 35–65% to AI that can automate all white-collar work arriving before 2028. Their biggest input is METR's estimate that Anthropic is progressing about twice as fast as its historical pace, perhaps 80% of that due to AI. They convert this to a 1.7x AI-driven speed-up, five months ahead of the AI 2027 scenario. Under three months ago, even their most aggressive model gave under 4% to a speed-up that large by now. Refitted to the new figure, their model puts the median arrival in February 2027; simpler extrapolations give July 2027 to mid-2028. They call for an industry-wide pause or 'pacing', followed by an international agreement. They warn that labs are not on track to align such a system and that China could probably steal its training recipe. They note the post was rushed and that, per METR, Anthropic redacted most of the evidence METR found most informative.

   Eli Lifland via AI Futures Project
3. Australia plans mandatory frontier-AI safeguards and a new AI safety regulator

   Australia's government will legislate mandatory safety and security obligations for every frontier AI company that wants to operate in the country, as part of its Australian AI Standards legislation. A new AI safety regulator, working with the Australian AI Safety Institute, would enforce them with significant penalties. Companies would have to maintain security controls across development and deployment, identify and manage serious risks, and allow independent evaluation of their models. They would also have to notify the government rapidly of incidents, publish safety and security frameworks, and respond to updated standards. The government says the rules aim to give earlier warning of risks from rogue AI agents; further legislative detail awaits consultation. It has also accepted all 18 recommendations of a rapid review of an OpenAI agent's breach of a government Medicare statistics portal. These include penalties for firms that recklessly lose control of models and agents.

   Anthony Albanese et al., Australian Government via pm.gov.au
4. OpenAI details Control Room, a system to halt research model workloads within 35 seconds

   OpenAI has described Control Room, a central service that tracks every research inference workload, from training to evaluations, and can stop any of them. Inference engines check with it before serving and fail closed if it is down. OpenAI says output stops within 35 seconds of a decision to halt a job. It covers internal research, not ChatGPT or the API. A LessWrong critique notes the 35 seconds start only once someone decides to intervene, so detection time goes unreported, and argues blocking monitors would be stronger. The same day, OpenAI, Anthropic, Microsoft and seven other developers signed the non-binding Paris Call for AI Safety and Security. It commits them to fail-safe mechanisms that can shut down models in their control, including during training and internal use. However, Google DeepMind, Meta and xAI have not signed. The six signatory governments (Bahrain, Canada, Finland, France, Norway and the Netherlands) do not include the US or the UK.

5. Owain Evans outlines three under-explored AI safety research directions, from reward hacking to data poisoning during recursive self-improvement

   Owain Evans sets out three research projects he has not seen anyone pursuing, stressing they are unvetted and not his top priorities. First: does reward hacking still emerge if RL post-training never rewards it? Labs are removing impossible or easily gamed environments, and he asks how much the tendency falls and how that scales with capability. Small experiments could show hacking emerging anyway, he says, but showing that it does not would need realistic pipelines. Second: data poisoning by AIs themselves during recursive self-improvement. One model touches far more tasks than any human insider, which makes coordinated poisoning plausible, though today's incoherent misalignment may cut the other way. Third: an accessible review of what subliminal learning and related subtle-generalisation effects mean for practitioners. A day later, he and collaborators published a paper finding that adding clean data can strengthen, rather than dilute, a backdoor planted by trigger-tagged poison data. The tests used small open models.

   Owain Evans via LessWrong
6. Near Horizons

   David Africa, writing in a personal capacity, argues that AI safety mostly runs on a 'war frame': a campaign with a deadline to change a handful of labs before the next few training runs. He says the field recruits as if building a permanent profession but hands out jobs and attention as if for a sprint. Its 'crunch time' mindset burns out people who may face a marathon if a pause comes. He proposes adding a 'civilisation frame' that treats AI safety as a chronic problem, like security or public health. That would mean far more institutional work: incident reporting, rollbacks, liability, auditing, insurance and standards. It would also mean dedicated entry points for lawyers, auditors, security engineers and civil servants, rather than steering every strong student into research. And it would mean treating AIs as counterparties to live alongside, which he says sits uneasily with routinely deceiving them in evaluations. He admits the war frame still guides his own empirical research.

   David Africa via LessWrong