On Monday November 2, 2026, Anthropic filed its first amendment to its public S-1 with the US Securities and Exchange Commission. The company that intends to list on Nasdaq this month under the ticker ANT used the Risk Factors section to tell investors that its internal estimates and expectations for future models "indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI.'"

That is a sentence about recursive self-improvement, written into a securities filing by a company that is about to go public. It is almost certainly the first time one has shown up there.

The specific sentence, and why it matters

TechCrunch's Julie Bort surfaced the passage in the S-1/A on Monday afternoon. The relevant text, from the filing's Risk Factors section:

"In addition, our increasing use of AI systems to accelerate our own research and development may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures. As a result, 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. For example, our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI.' We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data."

The admission inside the admission is "we have in the past not identified." Anthropic is telling investors — in the document that governs legal liability for post-IPO statements — that it has, before, failed to notice its own capability jumps until after they happened. And that it may not notice the next one until after it happens either. The RSP's whole architecture is built on being able to notice.

The filing uses the term the Loop audience will have been reading all year. From the same section:

"Using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress. This is one of the areas covered by our Responsible Scaling Policy, which sets out our plans to monitor and mitigate the risks as we approach what has been termed 'recursive self-improvement.'"

And the commitment, such as it is:

"We do 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, if and when that becomes technically possible. Accordingly, we are considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic."

What the Responsible Scaling Policy actually says about this

The RSP has gone through several rewrites. The current public version took effect in February 2026 after Time reported on an overhaul that dropped the earlier commitment not to train a system without guaranteed-in-advance safety measures. The threshold the S-1/A points to — "dramatically accelerated progress attributable to AI" — is one of the policy's capability triggers for additional safeguards and, in principle, slowdowns.

Fortune's coverage of a June 2026 Anthropic essay by two in-house researchers put the stakes in the words the company's own staff were using at the time: that AI systems may "autonomously design, build, and train their own successors," that "we are not there yet, and recursive self-improvement is not inevitable," and that "it could come sooner than most institutions are prepared for." Five months later, the same company is telling the SEC that "imminent" is now the right word for how close one of its policy triggers is.

Anthropic's own framing holds the two apart: current models have not been designated as crossing the threshold; internal estimates for future models suggest the threshold is close. That is a careful distinction, and it is also the exact distinction "we have in the past not identified" was offered to undermine. If the next capability jump is one the company only recognises after the fact — as it has in the past — then by the policy's design, the trigger could be met before the mitigation plan is ready to run.

The timing

The S-1/A sits a week ahead of the roadshow. Bloomberg's reporting, relayed via TheNextWeb in October, had the roadshow beginning the week of November 9 and the stock trading before the November 26 Thanksgiving close. The amendment is the standard pre-roadshow refresh — updated financials, cleaned-up risk disclosures, a few new paragraphs where the drafting committee has grown nervous since the October 26 initial S-1.

The Decoder ran the financial side of the amendment the same afternoon, citing Bloomberg and The Information: quarterly revenue near $17 billion, roughly twelve times the equivalent period a year earlier; operating costs higher than revenue; a net loss on the order of $23 billion for the quarter, against $12.1 billion a year prior. The Decoder notes that approximately "$5 billion for short-term compute capacity" locked in and "$2 billion in employee donation matching" are the two line items the company specifically named as contributors to the swing — not the full ~$11 billion gap on their own, but the ones the filing surfaces first. The New York Times's reporting, as summarised there, pegs Anthropic's compute and data-center lease commitments at $895 billion. That is a sentence that costs $895 billion.

The disclosure sits in the same amendment as the self-improvement warning. The structural news is not that these two disclosures happened to arrive together. It is that both appeared in the risk-disclosure document whose job is to say what could go wrong in a way that insulates the company from securities litigation afterwards.

What the filing is actually buying

A Risk Factors paragraph in an S-1/A does three things at once. It warns investors. It documents what the company believed at the time of the offering. And it defines the ground on which post-IPO lawsuits will be fought if things go badly. By writing "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 has pre-booked the defence to a specific class of future complaint: we told you we might miss it.

That is a defensible thing to do, and it is also an unusually frank thing to do. OpenAI's most recent public filings do not contain anything equivalent, and the frontier labs with no S-1 to file at all — Google DeepMind, xAI, Mistral — are not obliged to write any of this down. Anthropic has, in effect, used the mechanism of a public listing to put a recursive-self-improvement warning onto an SEC-indexed page the way other companies put product-liability warnings. The audience it is addressing is not just investors; it is the record.

The "considering updates" clause is where the policy work shows through. In practice it means an unspecified amount of further internal process before the RSP is amended again, with no stated end date and no stated commitment on what any revised version will contain. The optimistic read is that this is the company being honest that it does not yet know the answer. The cynical read is that "we are considering updates to our Responsible Scaling Policy" is a sentence anyone can write about anything at any time.

What to watch

1. The next RSP revision. The current public version dates to February 2026. The S-1/A commits Anthropic to "considering updates." The useful question is whether the revised policy keeps the trigger threshold where it is and strengthens the mitigation branch, or whether — as happened in the February rewrite — it quietly relaxes the trigger instead. Both would be defensible to a securities lawyer; only one is defensible to the person who wrote the September 2023 original.

2. Whether the SEC asks about it. The S-1/A is now in the staff review window. The Division of Corporation Finance writes comment letters on filings whose risk factors they find thin or confusing, and this risk factor is unusual enough — a self-reported "we may imminently cross our own policy threshold" paragraph is not a shape the staff has read before — that a comment letter would be plausible. If one lands, it becomes public twenty business days after the IPO closes, which is where the drafting-committee notes typically surface.

3. The IPO date. Bloomberg's reporting had the roadshow opening the week of November 9 and the stock trading before November 26. If either moves without a new amendment, the roadshow is going fine. If a second amendment lands first, something in the SEC-staff dialogue or the investor-day material produced a change important enough to re-paper the filing.

4. Whether another frontier lab writes a comparable paragraph. DeepMind's Shane Legg told the Financial Times in September 2026 that "we can't let capabilities get ahead of safety." OpenAI's own internal Misalignment Reporting Framework in September used the phrase "we do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer." Neither statement is in a securities filing. The question for the next twelve months is whether anyone else files one that is.

Anthropic has told the SEC that one of its own policy triggers for self-improving AI may be imminent, that it has a history of noticing capability jumps late, and that the policy under which it will respond is now itself under review. All three sentences are, in the specific way securities disclosure requires, true as of the filing date. All three sentences are also the kind of thing a company only writes down when the alternative — writing nothing — has become more expensive. The week in which the stock begins trading is the week in which the market gets to price that trade-off.