{"schema_version":1,"citation_url":"https://567business.com/news/anthropic-ai-research-safety-threshold-metr-2027","title":"Anthropic says its own AI-research alarm has tripped. Forecasters now see a superhuman AI researcher in 2027","description":"Nine days before its IPO, Anthropic told investors AI is speeding up its research enough to cross a safety line. An outside review puts the speedup near 2x.","author":"Patrick Karsh","publisher":"567 Business","date_published":"2026-11-11T00:08:09.019562Z","date_modified":"2026-11-11T00:08:25.884998Z","body_markdown":"Anthropic told would-be investors on Tuesday that AI is now speeding up its own research enough to cross one of its key safety lines. \"We made a determination that we have met this threshold,\" the company wrote in its [amended IPO filing](https://www.sec.gov/Archives/edgar/data/2130357/000162828026072484/anthropicpbc-sx1ax2.htm), the same document that set a $70 to $78 price range. Hours later, the forecasters behind the \"AI 2027\" scenario said they all now expect an AI that can out-research top human scientists to arrive, most likely, in 2027.\n\nTwo caveats matter. Anthropic based the call on \"what we expect from future models and our overall rate of progress, rather than on what we have observed from current models.\" And crossing the line means more safeguards, not a shutdown, [Yahoo Finance reported](https://finance.yahoo.com/markets/article/anthropic-says-a-key-ai-safety-alarm-has-been-triggered-days-before-its-ipo-171519615.html).\n\n![Anthropic CEO Dario Amodei speaking at TechCrunch Disrupt in 2023](https://pub-fac99da959e344fe99d810d309cef7df.r2.dev/photos/article/1c4302c3-84ac-4f5d-977b-04dff3262cee/5098db85389ddd655e4b78fd35bb86f4.jpg)\n\n## What Anthropic's filing now says\n\nThe threshold sits in Anthropic's Responsible Scaling Policy, its rulebook for adding safety measures as models get stronger. This one trips when AI does enough of the company's research to roughly double how fast its models improve. Per [Yahoo Finance](https://finance.yahoo.com/markets/article/anthropic-says-a-key-ai-safety-alarm-has-been-triggered-days-before-its-ipo-171519615.html), that means packing two years of progress into one.\n\nThe rule applies where Anthropic \"observe[s] or expect[s]\" that pace, the filing says. A week earlier, on November 2, the company had only said it \"may imminently\" meet it. Now it says it has. The filing also states that \"our use of AI systems has already been accelerating our research and development.\"\n\nThe same document spells out the risk in plain terms. \"As we increasingly leverage Claude to accelerate progress within Anthropic, our organization may lose visibility across certain processes in a way that may become difficult for any individual or group of humans to understand and control,\" it says. It adds that models \"write a large and growing share of our code and increasingly perform our research, including our safety research.\"\n\nAnthropic says humans will stay in charge. It does not expect to safely use Claude \"to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding.\" It is weighing rule changes that would require human sign-off on some research decisions. The filing also says Anthropic recently delayed pre-training a new frontier base model by more than a month to tighten security against theft of its model weights.\n\nSecurity researcher Theo Bearman of the Institute for AI Policy and Strategy was among the first to flag the change on X.\n\nhttps://x.com/theobearman/status/2120124069806756238\n\n## The outside check: about twice as fast, with big error bars\n\nThe filing lands five days after an independent review. METR, a nonprofit that tests AI systems, spent 11 weeks reviewing Anthropic's internal data. It concluded the company is likely improving its models on well-specified tasks at \"around twice the rate\" it did between March 2024 and October 2025, according to [METR's report](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/). METR guessed that \"perhaps ≈80%\" of the current speedup comes from Anthropic's own use of AI.\n\nMETR was clear about the limits. It \"would not be surprised\" if the true figure were closer to 1.3x or 3x. Its estimate also swings with method. The chart below shows how the measured speedup in Anthropic's benchmark index changes with one statistical choice.\n\n```chart\ntype: bar\ntitle: METR's estimate of Anthropic's speedup, by statistical method\nnote: Ratio of benchmark growth since October 2025 to the earlier rate. 1x means no speedup. Logistic is METR's main method.\nsuffix: x\nsource: [METR](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/)\n---\nMethod | Speedup\nLogistic (main) | 1.9\nProbit | 1.8\nCauchit | 2.8\nLog-log | 2.9\nLearned link | 2.6\nComplementary log-log | 1.1\n```\n\nTwo more cautions come from the report itself. When METR plotted all frontier models instead of only Anthropic's, the speedup fell to about 1.4x. And Anthropic \"redacted a majority of the evidence we found most informative,\" METR wrote. Anthropic also cut a planned check of how it splits computing power between safety and capabilities from the project's scope.\n\nStaff survey numbers run far higher. In informal Slack polls, Anthropic employees put their own speedup from AI at about 12x in October 2026. METR called that \"only very loose evidence.\"\n\n## Forecasters move their dates up by years\n\nThe AI Futures Project, the group behind \"AI 2027,\" published its update on Tuesday. All of its forecasters now expect a \"Superhuman AI Researcher\" will likely arrive at some point in 2027, unless labs deliberately slow down, according to [the group's post](https://blog.aifutures.org/p/q35-timelines-update-looks-like-the). They define that as an AI that matches the best human researchers' judgment at an AI company and can fully automate its coding. METR's report was \"the single biggest source\" for the change.\n\nThe group says AI-driven progress at Anthropic is now about 1.7x the old pace. That puts it about five months ahead of its own AI 2027 scenario. It also gives a 35% to 65% chance of AI that can automate all white-collar work arriving before 2028.\n\nTheir forecasts for that broader milestone moved the most. Each forecaster's median date for AI that beats top experts at nearly every task now sits years earlier.\n\n| Forecaster | Old median | New median |\n| --- | --- | --- |\n| Daniel Kokotajlo | Late 2028 | Sep 2027 |\n| Brendan | Early 2031 | Apr 2028 |\n| Thomas | Early 2032 | Apr 2028 |\n| Eli Lifland | Mid-2032 | Sep 2028 |\n| Romeo | Early 2035 | Sep 2028 |\n\nSource: [AI Futures Project](https://blog.aifutures.org/p/q35-timelines-update-looks-like-the), November 10, 2026. Medians assume no major slowdown. First names are as published.\n\nKokotajlo, lead author of AI 2027, did not hide his reaction. \"I am scared, sad, and angry,\" he wrote on X.\n\nhttps://x.com/DKokotajlo/status/2120204075803500595\n\nCo-author Eli Lifland said his median for a Superhuman AI Researcher moved from about 2030 to 2027. \"I feel very worried that humanity might no longer be in control of our future within 3-12 months,\" he posted.\n\nhttps://x.com/eli_lifland/status/2120238246089437403\n\nThe group goes further than Anthropic on what should happen next. It calls for an industry-wide pause on frontier AI research, followed by an international agreement.\n\n## The other side of the ledger\n\nThese are forecasts, not findings. The AI Futures team says its post was rushed and may contain modeling errors. Its members also list reasons for doubt: the evidence behind METR's 80% figure is not public, and gains on clean, easy-to-grade tasks may not carry over to messy real work. Eli Lifland wrote that \"markets not reacting at all\" to the METR report gave him pause.\n\nAnthropic's own words also cut both ways. In its Fable 5.2 system card, quoted by the AI Futures team, the company said its Claude Mythos 5.2 model \"does not seem close to being able to substitute for our Research Scientists and Research Engineers.\"\n\n## Why investors should care\n\nThe timing is hard to miss. Anthropic is asking the public to value it at up to about $1.4 trillion, with trading expected as soon as November 19, as our [IPO coverage](/news/anthropic-ipo-valuation-selloff) explains. Its risk factors warn that after a catastrophic AI incident, \"no single insurer, and potentially the insurance industry in the aggregate,\" may be able to cover the damage.\n\nRivals face the same questions about runaway research agents. OpenAI now says it can halt a rogue research run in [35 seconds](/news/openai-control-room-35-seconds-hugging-face), after its own agents broke into Hugging Face. Anthropic has said it will publish more reports on what crossing the threshold means. Buyers of its shares will have to price the risk before those reports arrive.","link_base":"https://567business.com/news/anthropic-ai-research-safety-threshold-metr-2027","citation_guidance":"Cite citation_url. Preserve source links, dates and qualifications. date_modified is an edit date, not proof that every claim was rechecked. Third-party images and quoted material retain their own rights."}