Anthropic told investors in an amended IPO filing on Monday that its own use of AI to build AI could make progress speed up faster than its safety work can follow, and that it “may imminently” meet its Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.” The company says it has not labeled any current model as past that threshold and that it is weighing changes to keep humans in control of frontier development. The language sits in the risk-factors section of the revised S-1/A, as reported by TechCrunch.

What the filing says

According to TechCrunch’s reading of the document, Anthropic acknowledges that it already uses AI systems inside its research and development. It then adds a warning: 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 points to the Responsible Scaling Policy, the company’s published framework for monitoring and limiting risks as models get more capable, and ties the concern to what the industry calls recursive self-improvement, or RSI.

Two statements are new in this version. The first is that, because of its growing use of AI in research, the company has in the past failed to recognize a significant acceleration in its development until after it happened, and may do so again. The second is that internal estimates and expectations about future models, not its current ones, indicate it may imminently reach the policy’s acceleration definition. The filing adds that the company continues to assess the underlying data.

What Anthropic says it will not do

The same passage sets a limit. Anthropic states that it does not expect to be able to safely use Claude to design and build successive generations of frontier models fully autonomously, without meaningful human supervision or understanding, “if and when that becomes technically possible.” It says it is considering updates to the Responsible Scaling Policy to keep humans in control of frontier model development. No details of those updates were given in the coverage we reviewed, so the practical effect is unknown until the policy itself changes.

The financial backdrop

The risk language arrived alongside updated financials. The Decoder, citing Bloomberg and The Information, reports quarterly revenue of nearly $17 billion, about twelve times a year earlier, against operating costs of $17.5 billion and a net loss of $23.3 billion for the quarter. Anthropic attributed the cost jump to roughly $5 billion of short-term compute capacity and $2 billion of employee donation matching. The updated prospectus is also reported to list $895 billion in compute and data-center commitments, and the offering could raise around $100 billion according to the New York Times as relayed by The Decoder. We have not seen the filing text ourselves, so these figures are as reported by those outlets.

For earlier coverage of how the market may price the offering, see our report on Anthropic’s expected credit rating after the IPO.

How to read it

An S-1 is a legal disclosure, and companies include warnings so they cannot be accused of hiding risks. TechCrunch notes that the document also works as marketing, since a claim of being close to systems that can improve themselves signals capability to investors. Both readings can be true at once. What is different here is the specificity: the company names its own policy threshold and says it may reach it soon, rather than speaking about RSI in general terms.

Self-directed research is no longer only a theoretical worry. Last month we covered Recursive’s Flint, a coding model largely built by its own research system, and DeepMind’s Levels of AGI framework gives one vocabulary for describing how capable such systems are. Neither establishes that a runaway loop is near, and the filing itself says none of Anthropic’s models has crossed its line.

What it means for regulators and customers

Disclosure of this kind gives lawmakers and enterprise buyers a dated, company-signed statement to point to. Securities law makes the risk factors something the company can be held to, which is a different pressure than a blog post. Customers that rely on Claude for coding, math or security work should watch for any change to the Responsible Scaling Policy, because it governs what the company will release and under what safeguards.

What to watch next

  • Whether Anthropic publishes revisions to its Responsible Scaling Policy and what “keeping humans in control” means in practice.
  • Whether other labs preparing public offerings add similar language to their own risk disclosures.
  • The next amendment to the S-1 and the pricing of the offering.
  • Independent evaluations, such as audits by outside safety groups, of how much of Anthropic’s research is now done by AI.
  • Responses from US and EU regulators to a company describing its own acceleration risk.

Key takeaways

  • Anthropic’s amended S-1 says it may imminently meet its policy definition of dramatically accelerated AI-driven progress.
  • No current model has been designated as exceeding that threshold, according to the filing.
  • The company says it does not expect to safely hand frontier model design to Claude without meaningful human oversight.
  • Reported quarterly revenue was nearly $17 billion with a $23.3 billion net loss, per Bloomberg and The Information as relayed by The Decoder.
  • An S-1 is both a legal disclosure and a sales document, so the statement should be read with both purposes in mind.

Frequently asked questions

What is recursive self-improvement?

It is the idea that an AI system becomes good enough at AI research to build better versions of itself, which then do the same, speeding up progress with each generation. Anthropic’s filing treats it as both a source of benefit and a safety risk.

Has Anthropic built a self-improving AI?

No. The filing says none of its current models has been designated as exceeding its acceleration threshold. It does say the company uses AI in its own research and may reach the threshold soon.

Where does the warning appear?

It appears in the risk-factors section of Anthropic’s revised S-1/A filed on Monday, as reported by TechCrunch. We relied on that reporting and did not review the full filing.