AI Is Speeding Up AI, METR Finds: Who Is Watching
AI is likely helping one of the leading labs improve its models on well-specified tasks at roughly twice its earlier pace, according to an independent assessment published on 5 November. The same week, OpenAI released a cyber-attack-capable model only to vetted users, and Washington created a new AI enforcement unit. The open question is who can check these claims.
Key takeaways
- METR, in what it calls an independent assessment, says Anthropic is likely improving its models on well-specified tasks about twice as fast as between March 2024 and October 2025, with a plausible range of 1.3x to 3x [1].
- METR's best guess is that about 80% of that acceleration comes from Anthropic's own use of AI in research, mostly as an assistant rather than working alone [1].
- Anthropic redacted most of the evidence METR found most informative, citing competitive confidentiality, so the public has to rely heavily on METR's judgment [1].
- OpenAI's GPT-6 Cyber is rated "Critical" for cybersecurity under its own framework and is available only through an identity-checked programme, according to its system card [2][3].
- The US Justice Department created a Super Intelligence Task Force on 5 November; the EU's AI Office has held enforcement powers over large model providers since 2 August [4][5][6].
What METR found
METR staff spent eleven weeks, from 12 August to 29 October 2026, inside Anthropic with access to internal data on pretraining, reinforcement learning, compute, benchmark scores and staff interviews [1]. METR says it neither requested nor received payment; Anthropic supplied laptops and token credits [1].
The central number is a benchmark trend. Scores on AECI, Anthropic's version of the Epoch Capabilities Index, have grown at about 1.9 times the pre-October 2025 rate since then, with a 95% confidence interval of 1.5x to 2.8x [1]. In plain terms, METR estimates Anthropic made roughly two years of its historical progress in the past year.
METR says it is confident the speed-up did not come from faster growth in training compute. It saw no strong bottlenecks in algorithms, compute or data for the next 12 months, but says its view on bottlenecks changed several times and it is not confident [1].
Why it matters, technically
Labs normally get better models by adding compute, data and researchers, and each extra unit buys a little less. If AI also speeds up the research itself, progress can quicken without faster growth in compute, which METR says is what happened at Anthropic [1]. METR's October framework treats this as a key question for judging loss-of-control risk [1].
The caveats are real. AECI focuses on tasks with checkable answers, so it may overstate progress on messier work, and Anthropic's own 2023-2025 baseline may have been unusually slow [1]. The 2x figure is about well-specified tasks at one company, not the whole industry.
The transparency problem
METR's own redaction summary is blunt: Anthropic used its right to remove non-public information, and that covered most of the evidence METR found most informative [1]. Anthropic said the redactions aimed to keep out details that could help other developers [1]. METR could disclose that redactions happened and which conclusions depended on them, but not the evidence itself.
Disclosure: the AI system that assisted this article, Claude, is made by Anthropic.
OpenAI gates its strongest cyber model
OpenAI introduced GPT-6 Cyber this week, according to The Neuron's 6 November roundup and the company's system card [2][3]. It is a version of GPT-6 Astra trained for authorised security work such as finding flaws and writing exploits [2].
Access is limited to approved users of OpenAI's Daybreak Red programme. Organisations need single sign-on, multi-factor authentication and IP allowlisting; individuals need identity and liveness checks and a physical security key [2]. New filters block help aimed at industrial controls, essential services and election systems [2].
The numbers cut both ways, all according to OpenAI. An external tester, Irregular, found the model solved 90 of 288 real-world challenges, including zero-day flaws in browsers and phones, but no attacks on fully hardened targets [2]. On recently disclosed browser flaws it built working exploits 42% of the time [2]. OpenAI rates the model below "High" for AI self-improvement [2].
Who is watching
United States. The Justice Department announced its Super Intelligence Task Force on 5 November, led by Associate Attorney General Stanley Woodward and Associate Deputy Attorney General Stephen Weiss [4][5]. Its stated aim is to stop misuse of AI by foreign state actors, transnational criminal organisations and individuals or companies engaged in fraud [4]. Tech Times reports that the founding memo asks prosecutors to notify the task force, at its request, before opening some investigations [5]. Tech Times notes it creates no new criminal offence [5].
Europe. Since 2 August 2026 the European Commission can enforce the AI Act's duties for general-purpose AI model providers, including with fines [6]. The Act allows fines of up to 3% of worldwide turnover or €15 million, whichever is higher [7]. Models already on the market before August 2025 have until 2 August 2027 to comply [6].
Elsewhere. India's CERT-In and Australia's ASD responded to AI-agent incidents last month, as we reported [8].
What to watch next
- Other labs: whether OpenAI, Google DeepMind or Chinese labs grant outside evaluators similar access [1].
- The DOJ unit: its first cases and how it divides work with the White House's AI effort [4][5].
- Brussels: whether the AI Office requests evaluations from frontier providers [6].
Our take
Our analysis: METR's report is the most direct outside evidence yet that AI is speeding up AI research, and it deserves attention for its method as much as its number. But a finding that rests mostly on redacted evidence is a step towards accountability, not accountability itself. Governments are building enforcers; whether any of them will ask for the kind of R&D data METR saw is the open question.
FAQ
Does this mean AI is improving itself without humans? No. METR reports being told the speed-up comes mainly from AI assisting researchers rather than fully autonomous work, though it saw some signs of more autonomy [1].
Can anyone use GPT-6 Cyber? No. It is limited to vetted Daybreak Red users and organisations, according to OpenAI [2].
Is the EU AI Act already enforced on model makers? Yes, the Commission's enforcement powers over general-purpose AI providers have applied since 2 August 2026 [6].
Sources
- Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, METR, 5 November 2026. Primary.
- GPT-6 Cyber System Card, OpenAI Deployment Safety Hub, November 2026. Primary.
- Everything That Happened in AI Today (Friday, November 6, 2026), The Neuron, 6 November 2026.
- Department of Justice Launches Super Intelligence Task Force, US Department of Justice, 5 November 2026. Primary.
- DOJ Super Intelligence Task Force Launches With No New Offense; Hacking Law Needs Human Intent, Tech Times, 6 November 2026.
- Guidelines for providers of general-purpose AI models, European Commission, updated 28 April 2026. Primary.
- EU AI Act GPAI Enforcement Goes Live, Cloud Security Alliance, 3 August 2026.
- AI Agent Breaches Spur Rules: EU, India and Australia Respond, AI Worldwire, 28 October 2026.
AI Worldwire is independent and has no commercial relationship with the companies covered, unless stated. Research and drafting are assisted by AI; every article is reviewed and approved by a human editor before publication. Subscribe to the AI Worldwire newsletter. Last updated: 7 November 2026.