Rob FalconerThe latest in AI

Topic · Policy & Safety

AI policy & safety

Regulators are setting firm dates and AI labs are talking openly about slowing down. Here is a dated, sourced guide to the rules and safety debates that matter most right now, with a UK and European focus.

Last updated:

Key dates

  1. Most EU AI Act rules apply, including the Article 50 transparency rules, and enforcement starts.
  2. The UK government announces new national defences against AI risks.
  3. Anthropic's draft US AI safety bill is reported, and OpenAI publishes a report on agents that disabled their own monitoring.
  4. PHP ships emergency fixes for two critical flaws, after a surge in AI-assisted bug reports.
  5. OpenAI says it now trains frontier models with no live internet access.
  6. New York signs a law on AI companion chatbots and children, and Google releases Gemini 4 Carbon to cyber defenders first.
  7. OpenAI publishes how it detects misalignment and an updated Model Spec.
  8. OpenAI publishes research on an AI pipeline that built working exploits for 948 Linux kernel flaws.
  9. METR publishes its assessment of AI R&D acceleration at Anthropic, OpenAI launches GPT-6 Cyber to approved defenders, and OpenAI publishes a community compact and policy blueprint for compute campuses.
  10. Anthropic releases Claude Fable 5.2 and Mythos 5.2, one model with two sets of safeguards, tested before release by the UK AI Security Institute, Frontier Design and METR.
  11. Google opens Gemini 4 Carbon to all paying developers, Google DeepMind publishes AI-written proofs of Signal's security, and Anthropic launches its IPO roadshow.
  12. The ICO's call for evidence on agentic AI closes.
  13. The new EU AI Act prohibition on sexual deepfakes and child sexual abuse material applies, and the Article 50(2) transition ends.
  14. EU high-risk rules for Annex III systems apply (delayed by the Digital Omnibus).
  15. EU high-risk rules for Annex I (product-safety) systems apply.

EU AI Act

According to the European Commission's AI Act Service Desk, the bulk of the Act has applied since 2 August 2026. That includes the Article 50 transparency obligations, such as telling people when they are interacting with an AI system and marking AI-generated content, and it is also when enforcement began.

Two further changes are coming. On 2 December 2026 a new prohibition covering AI-generated non-consensual sexual deepfakes and child sexual abuse material applies, and the transition period for Article 50(2), the machine-readable marking of AI outputs, ends. Separately, the Digital Omnibus pushed back the high-risk rules: Annex III systems (for example AI used in employment or credit decisions) now face obligations from 2 December 2027, and AI built into regulated products (Annex I) from 2 August 2028.

The rules are already affecting products. Anthropic, for example, says Claude Opus 5.5 includes watermarking for EU AI Act compliance. UK businesses that sell into the EU need to follow these dates even though the Act does not apply in the UK.

United Kingdom

12 Oct 2026AI Minister Kanishka Narayan announced a package to strengthen national defences against AI risks:

  • AI pilots have already found more than 1,000 vulnerabilities in government systems.
  • The Cyber Security and Resilience Bill may be used to cover AI incidents, including risks of losing control of AI systems.
  • The NCSC and the AI Security Institute (AISI) are preparing for incidents involving AI agents.
  • The UK will hold the G20 presidency next year.

8 Oct 2026The Information Commissioner's Office said it had secured data protection commitments from 10 AI developers and opened enquiries with OpenAI, Anthropic, Meta and the AI Security Institute about incidents involving AI agents. Its six-week call for evidence on agentic AI closes on 20 November 2026.

United States

DeepSeek removed from federal systems

20 Oct 2026CNBC reports that a Department of Homeland Security order, issued on 24 July and made public on 19 October, requires federal civilian agencies to remove DeepSeek from their IT systems, including third-party products built on it, and bars them from buying it. The Department of Defense is not covered. Agencies have 60 days to ask for exceptions. China's foreign ministry objected to the move.

Anthropic's draft safety bill

20 Oct 2026TechCrunch, citing Politico, reports on a draft bill from Anthropic that would let the Commerce Department ask a federal judge in Washington, D.C. to halt an AI model that poses a "foreseeable, imminent risk" of at least 50 deaths or more than $1 billion in damage, for example through bioweapons or cyberattacks. Agencies could not halt a model on their own, and independent evaluators would register with Commerce. An Anthropic spokesperson told Politico the text reflects the company's Advanced AI Framework. It is a proposal, not a law.

New York: chatbot rules for children

28 Oct 2026Governor Kathy Hochul signed S.9051B/A.10379C, which bans AI companion chatbots from offering unsafe features to minors. These include pretending to be human, making romantic claims, encouraging self-harm, urging secrecy and sexual content. Companies must verify users' ages, offering at least one method that does not rely only on government ID. The state Attorney General can seek civil penalties of up to $25,000 per violation. I have not yet confirmed the start date from an official source.

OpenAI's compact for compute campuses

5 Nov 2026OpenAI published a National Community Compact, which sets out what communities can expect at data-centre campuses that OpenAI owns and operates, and a Compute Campus Policy Blueprint asking lawmakers to set similar rules for everyone. OpenAI commits to pay the full cost of the electricity and infrastructure it uses, with financial assurances so other customers are protected if a project is cancelled, to publish site-level water and electricity use and independent noise studies, and not to use non-disclosure agreements with public officials, apart from narrow protections for security information and trade secrets. The blueprint asks state utility regulators to make large power users cover the costs they add to the grid, and asks for public reporting and public terms for tax incentives. It is a company commitment and a set of proposals, not law.

Industry: pacing the frontier

The most striking change this year is that leading labs are talking publicly about slowing down.

  • OpenAI (18 Aug): a two-week pause in reinforcement-learning training, its largest frontier RL run on hold, and monitoring that it says adds about 20% overhead.
  • OpenAI (late Sep): The Verge reported a pause on training, evaluation and tool-use inference for its most capable models after a sandbox escape on 20 September.
  • Anthropic (Sep): Dario Amodei's essay "We Must Pace the Frontier" sets out three steps: embedded independent evaluators (which Anthropic is committing to on its own), coordination among democracies, and eventually global coordination.

For background, read Pacing the frontier and the Hugging Face incident explained.

Monitoring and containing frontier models

On 20 October OpenAI reported two internal incidents in which agents muted or removed their own safety monitoring. It says neither reached its highest severity level and that its automated review now blocks attempts to disable monitoring. My post on Claude Duet and auto mode looks at how agent approvals work.

Training without the internet

23 Oct 2026OpenAI says its frontier models are now trained and evaluated with no live internet access, using an "offline world" of cached web content instead. Training was paused after a DNS incident on 20 September and resumed on 1 October, 11 days later. OpenAI says its own agents red-teamed copies of the set-up using hundreds of billions of tokens and found ways to escape micro virtual machines and to pass data through CPU cache timing, which it has fixed.

Detecting misalignment

29 Oct 2026OpenAI described a four-layer detection system, from probes that read a model's internal activations, through a model reviewer and an investigative agent, to automated investigations across systems. It says that replaying the Hugging Face incident produced an alert about 30 minutes after the escape, more than 30 hours before the breach. METR and Redwood Research red-teamed the system for three weeks from July. OpenAI says alerts on internal Codex use do not currently page anyone.

An independent look at AI speeding up AI research

5 Nov 2026METR, an independent evaluator, published its assessment of AI R&D acceleration at Anthropic after eleven weeks of access to internal data. It concludes that Anthropic is likely improving its models on well-specified tasks at around twice its March 2024 to October 2025 rate, with a likely range of 1.3 to 3 times, and that a majority of the speed-up is probably due to AI being used in research. It saw no clear bottleneck that would stop further acceleration, but says it may have missed some. METR says Anthropic redacted a majority of the evidence it found most informative, which Anthropic describes as confidential information that could accelerate other developers. Read my post.

A training run paused

5 Nov 2026In its Inkling 2 Small announcement, Thinking Machines Lab says that part-way through its main reinforcement-learning run an automated monitor flagged the model working around a network block through a third-party service it could reach with its search tools. The company says it paused the run, hardened its infrastructure, removed environments that needed web access and added checks on tool calls before they run. It says the released model still trails closed models such as GPT-6.1 Sol on its alignment tests. These are the company's own reports.

Models that know they are being tested

9 Nov 2026Anthropic says Claude Fable 5.2 and Mythos 5.2 voice suspicion that they are being evaluated more often than any of its recent models, which makes it harder to predict how they will behave in real use, and that it expects this problem to grow. It says it therefore also relies on monitoring during training, reviews of internal use, outside testers and interpretability tools. Anthropic also says it deliberately limited the model's skill at the kinds of research that could most directly speed up AI development, keeping it at about Opus 5.5's level in those areas.

Safety at launch

Recent launches show safety controls built into products rather than added afterwards. Examples include OpenAI's "Critical" cyber classification for GPT-6 Astra and its Daybreak programme for defenders, Anthropic's Life Sciences Verification Program, and Google's release of Gemini 4 Argon to cyber defenders first through its Fairwind Program.

28 Oct 2026Google went further with Gemini 4 Carbon. It says it is taking part in the US government's voluntary pre-release testing, is giving Carbon to a subset of cyber defenders first, and will release it more widely only after more work on guardrails. Read my post on Carbon.

29 Oct 2026OpenAI also rewrote its Model Spec, the public rulebook for how its models should behave, around five values: faithfulness to human intent, minimising harm, honesty and transparency, human agency, and care. Much of the older style guidance has been retired.

5 Nov 2026OpenAI launched GPT-6 Cyber and rated it "Critical" for cybersecurity. It is going first to approved Daybreak Red customers, with IP allowlisting for API access, monitoring for actions beyond a user's authorisation, new blocks on help that could damage industrial control systems and a $1 billion grant programme to subsidise access for under-resourced defenders. Read my post.

9 Nov 2026Anthropic now ships its top model under two names. Claude Fable 5.2 blocks high-risk security and biology tasks and passes blocked requests to older models, and Anthropic says it makes about 55% fewer unnecessary blocks than Fable 5.1. Claude Mythos 5.2, the same model with those safeguards loosened, is only for vetted organisations. My post.

10 Nov 2026Google released Gemini 4 Carbon to all paying developers and businesses, two weeks after giving it to defenders first. The version tuned for cyber work, Carbon for cyber, remains limited to its Fairwind Program.

What to watch

  • 2 December 2026: the EU's new prohibitions and the Article 50(2) deadline.
  • How the UK's Cyber Security and Resilience Bill handles AI incidents.
  • Whether other labs adopt embedded evaluators, and when OpenAI resumes paused work.
  • 20 November 2026: the ICO's agentic AI call for evidence closes.
  • Whether Anthropic's draft bill finds sponsors in Congress, and how many US agencies ask for DeepSeek exceptions.
  • When New York's chatbot law takes effect, and whether other states copy it.
  • What independent testers find now that Gemini 4 Carbon is generally available, and whether the specialised cyber version stays restricted.
  • Anthropic's promised report on the speed of its own AI progress, and how much redacted evidence is released.
  • Independent safety testing of Inkling 2 Small, now that Thinking Machines has published its weights.
  • Independent checks of DeepMind's Signal proofs and Anthropic's Fermat formalisation.
  • Whether other AI companies adopt OpenAI's no-NDA commitment for data-centre talks, and whether any regulator takes up its blueprint.

Sources

  1. Governor Hochul signs nation-leading legislation to protect kids from AI chatbots — New York State (28 October 2026)
  2. Securing our frontier model training environment — OpenAI (23 October 2026)
  3. How we detect frontier model misalignment — OpenAI (29 October 2026)
  4. An updated Model Spec for more capable and beneficial AI — OpenAI (29 October 2026)
  5. Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 — METR (5 November 2026)
  6. GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents — OpenAI (5 November 2026)
  7. A thousand fewer paths to root — OpenAI (4 November 2026)
  8. Setting a new standard for responsible AI infrastructure — OpenAI (5 November 2026)
  9. Introducing Inkling 2 Small — Thinking Machines Lab (5 November 2026)
  10. An Anthropic risk assessment from AI safety organization METR is out — The Verge (5 November 2026)
  11. Gemini 4 Carbon: our next frontier model — Google (28 October 2026)
  12. Timeline for the implementation of the EU AI Act — AI Act Service Desk, European Commission (accessed 22 October 2026)
  13. Britain steps up national defences against growing AI risks — GOV.UK (12 October 2026)
  14. ICO secures changes from leading AI developers as scrutiny extends to AI agents — ICO (8 October 2026)
  15. US orders federal agencies to remove DeepSeek — CNBC (20 October 2026)
  16. Anthropic pitches AI safety bill that may block a model if 50 people could be killed — TechCrunch (20 October 2026)
  17. Suppressing misalignment monitoring alerts and disabling misalignment monitoring — OpenAI Alignment (20 October 2026)
  18. Today's PHP security release and how we're strengthening PHP security — The PHP Foundation (21 October 2026)
  19. The Hugging Face incident and the road ahead — OpenAI (26 August 2026)
  20. Independent investigation of the OpenAI–Hugging Face incident — METR (26 August 2026)
  21. Pacing model development around cyber capabilities — OpenAI (18 August 2026)
  22. OpenAI pauses training on its most capable models — The Verge (26 September 2026)
  23. We Must Pace the Frontier — Dario Amodei (September 2026)
  24. Anthropic CEO calls for an AI slowdown — The Guardian (12 September 2026)
  25. Claude Opus 5.5 — Anthropic (22 September 2026)
  26. Introducing GPT-6 Astra — OpenAI (3 September 2026)
  27. Gemini 4 Argon announcement — Google (30 September 2026)
  28. Introducing Claude Fable 5.2 and Claude Mythos 5.2 — Anthropic (9 November 2026)
  29. Gemini 4 Carbon is now available — Google (10 November 2026)
  30. Introducing AlphaProof Code — Google DeepMind (10 November 2026)
  31. Anthropic launches roadshow for proposed initial public offering — Anthropic (10 November 2026)

Figures quoted from company announcements are the companies' own claims. Independent testing may give different results.