Anthropic amended its IPO filing Monday with the Securities and Exchange Commission1. The filing warns the company may soon cross a safety threshold called recursive self-improvement. Crossing it means AI designs a smarter version of itself, with little work from a human in the loop. Anthropic named this same risk five months earlier, in its own research7.
The Threshold Anthropic Put a Name On
Anthropic's own researchers described the risk first. "We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for," the Anthropic Institute wrote in June7. Five months later, its lawyers put a number on that guess.
The amended filing defines the threshold as a doubling of Anthropic's own pace of progress, with AI responsible for the gain. "Using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress," the filing states1. Anthropic keeps every current model under that line, by its own account.
A blog post and a legal filing read differently to a regulator. Lying in the second carries legal liability; lying in the first costs only reputation. That gap is the whole reason this warning lands louder than the one Anthropic wrote in June.
The City Council Heard It First
New York City Council held a hearing on AI risk Oct. 5, a month before the filing6. Jacob Coxon testified first. He worked at OpenAI and at Anthropic before he left both labs.
"The companies run on a startup mindset: move fast, break things, fix them later. That works for a photo sharing app. It does not work for building the most powerful technology ever built," Coxon told the council4.
Alex Turner spoke under subpoena the same day. He left Google DeepMind's alignment team before he testified. "I myself would guess AI takeover chances at roughly one in three. This logic may sound sci-fi, but it is a mainstream scientific concern that has been discussed for many years," he said5.
Both witnesses named the risk Anthropic's filing later put a number on. Their testimony landed 28 days before the amended filing, person first and lawyers second. Coxon and Turner spoke in a council chamber, carrying only their own names and former jobs behind their words.
Whistleblowers Named the Risk 28 Days Before Anthropic's Filing Did
Dated testimony, essays and filings on AI self-improvement, June 4 to Nov. 3, 2026
Each point marks one dated statement. The two City Council dots land a month before the amended filing, with Kokotajlo's reaction landing the day after.
Sources: TechCrunch [1]; Daniel Kokotajlo on X [2]; Yahoo News [4]; Yahoo News [5]; The Anthropic Institute [7]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Date | Event |
|---|---|
| June 4, 2026 | Anthropic essay. Says self-improvement 'could come sooner than most institutions are prepared for' |
| Oct. 5, 2026 | Coxon testifies. Calls the industry's pace 'extremely reckless' |
| Oct. 5, 2026 | Turner testifies. Puts AI takeover odds at roughly one in three |
| Nov. 2, 2026 | Amended filing. Names the RSI threshold in a legal disclosure |
| Nov. 3, 2026 | Kokotajlo reacts. Calls the timeline to superintelligence short |
The Skeptic's Ledger
Ross Hendricks reads Anthropic's warnings on a different ledger. He is a Porter and Company analyst who ties the company's biggest warnings to its biggest financial pressure. "Anyone still wondering why the company is threatening to end humanity unless the government shuts down their competition?" he asked in September, after an earlier leak of Anthropic's numbers3.
Ross Hendricks@Ross__HendricksReuters just released a sneak peak of Anthropic financials, showing the company reported a net loss of $42 billion in 2025 - and this is before it started losing market share to dirt cheap open source models
Anyone still wondering why the company is threatening to end humanity…
Ed Zitron makes a related case about money, from a newsletter that tracks tech spending closely. He pegged Anthropic's IPO near $1.4 trillion, below the company's earlier target, and called the listing "a cynical attempt to use retail investors as exit liquidity"8.
Each critic accepts the technical definition Anthropic wrote. Their dispute is about the month the company chose to publish it.
Anthropic built a structure meant to keep that pressure at arm's length. Chief Executive Dario Amodei and six other co-founders hold 50.1 percent of the company's voting power through an entity called Founder LLC10. The arrangement lets them outvote every outside shareholder on board seats and other matters. Anthropic says the buffer protects its safety mission from the market Hendricks and Zitron describe.
Seven Co-Founders Hold 50.1 Percent of Anthropic's Vote
Share of Anthropic's voting power held through Founder LLC against every other shareholder combined
The ring is all of Anthropic's voting power. The magenta arc is the 50.1 percent seven co-founders control through Founder LLC, and the gray arc is every other shareholder's combined vote.
Source: CNBC [10]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Founder LLC's seven co-founders | 50.1% |
| Every other shareholder | 49.9% |
| Whole | 100% |
What Anthropic Itself Admits
The filing carries its own limit. "We have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred," Anthropic wrote1.
Daniel Kokotajlo read that limit and still took the filing seriously. He leads the AI Futures Project, the research group behind the AI 2027 forecast. "It seems timelines to full RSI and superintelligence may be very short indeed," he wrote on X, hours after the filing became public2.
Anthropic set a second limit on itself too. The company wrote that it does "not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding"1. That sentence works as a promise and a confession together. Anthropic promises human oversight, and admits oversight might arrive too late to catch the moment that matters.
What to do
Read Anthropic's filing as a floor, apart from a verdict. The company named a threshold its models have stayed under so far, by its own account. That gives regulators and investors a fixed point to measure the next model against. The number carries more weight once it sits inside a document filed under oath.
Watch for an independent group to test a future Anthropic model against the Responsible Scaling Policy's own definition. Anthropic wrote that it might miss the acceleration until afterward. An outside check matters more than the company's word alone.
By the numbers
- $1.4 trillion: low end of the valuation range tied to Anthropic's amended IPO filing8.
- Five months: the gap between Anthropic's own essay on self-improvement and the same warning inside a legal filing7.
- Oct. 5: the date New York City Council heard whistleblower testimony on AI control, a month before the filing6.
- One in three: Alex Turner's own estimate of the odds of an AI takeover, given under subpoena5.
- 100 percent: the share of Anthropic's current models that stay under its own threshold, by the company's account1.
- Nov. 3: the date Daniel Kokotajlo called the timeline to superintelligence short, one day after the filing became public2.
What to watch
Watch whether Anthropic names a model against its own threshold before its IPO roadshow ends this month. A confirmed crossing would move this story from a filing's risk factor to a live event every investor has to price.
Congress is worth watching too. Rep. Ro Khanna's bill to ban self-improving AI sat with few co-sponsors in September. A filing this specific could change that math fast9.
Sources
- "Anthropic warns that it's very close to RSI capability, aka self-building AI," TechCrunch, Nov. 2, 2026, https://techcrunch.com/2026/11/02/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai/
- Daniel Kokotajlo, "Post on X," X, Nov. 3, 2026, https://x.com/DKokotajlo/status/2117738240757371022
- Ross Hendricks, "Post on X," X, Sept. 28, 2026, https://x.com/Ross__Hendricks/status/2104717983566176452
- Taylor Herzlich and Hannah Fierick, "Anthropic whistleblower Jacob Coxon doubles down on AI warnings at NYC hearing: 'Extremely reckless'," Yahoo News, Oct. 5, 2026, https://www.yahoo.com/news/politics/articles/anthropic-whistleblower-jacob-coxon-doubles-145451543.html
- "Whistleblower says AI takeover is a 'mainstream' scientific concern," Yahoo News, Oct. 5, 2026, https://www.yahoo.com/news/science/articles/whistleblower-says-ai-takeover-mainstream-195319892.html
- "New York City Council Convenes Hearing on Artificial Intelligence Risks with Testimony from AI Executives, Whistleblowers, Experts," New York City Council, Oct. 5, 2026, https://council.nyc.gov/press/2026/10/05/3278/
- Marina Favaro and Jack Clark, "When AI builds itself," The Anthropic Institute, June 4, 2026, https://www.anthropic.com/institute/recursive-self-improvement
- Ed Zitron, "Post on X," X, Oct. 28, 2026, https://x.com/edzitron/status/2115504670722474389
- Garrett Downs, "Khanna to introduce AI safety bill with ban on 'recursive' technology until safeguards exist," CNBC, Sept. 28, 2026, https://www.cnbc.com/2026/09/28/khanna-ai-safety-bill.html
- Ashley Capoot, "Anthropic IPO filing gives 7 founders 50.1% voting control," CNBC, Oct. 26, 2026, https://www.cnbc.com/2026/10/26/anthropic-ipo-founder-voting-control.html
