Anthropic, the company behind Claude, has told investors that it may soon cross one of its own safety red lines: the point where AI itself dramatically speeds up the building of AI. The warning sits in the risk factors of the company’s amended IPO prospectus (S-1/A, filed with the SEC) on Monday, 2 November 2026, and was first spotted by TechCrunch.
What the filing says
The key passage is new. According to Gizmodo, the version of the filing from 26 October did not contain it. It reads:
“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 Responsible Scaling Policy (RSP) is Anthropic’s self-imposed rulebook that ties stricter safety measures to specific capability levels. The filing makes clear which risk this threshold is about: “Using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress,” it says, adding that the RSP sets out plans to monitor and mitigate the risks “as we approach what has been termed ‘recursive self-improvement.'”
Perhaps the most striking line is an admission about the past. Anthropic writes that its growing use of AI in its own research may let capabilities “advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures,” and that “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.”
What “recursive self-improvement” means here
Recursive self-improvement, often shortened to RSI, describes AI that gets good enough at AI research to build its own, better successor, which then builds the next one faster still. Some see it as the gateway to a scientific boom; others as the moment AI could slip out of human control.
Anthropic’s filing does not claim that this has happened. The RSP threshold it mentions is an earlier, narrower milestone: progress that is dramatically accelerated by AI, while humans still steer. Elsewhere in the prospectus, the company describes how Claude already takes part in building its successors, “running experiments, generating and processing training data, automating evaluations and safety audits, and writing and reviewing code.” Anthropic calls this dynamic “closing the loop.”
Humans stay in the loop, for now
The company also draws a line it says it won’t cross: “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.” It adds that it is considering updates to its Responsible Scaling Policy “to keep humans in control of frontier model development at Anthropic.”
The context matters too. The same filing shows how fast Anthropic is growing: revenue went from zero at Claude’s commercial launch in early 2023 to $17.0 billion for the three months ended 30 September 2026, according to the prospectus. Risk factors in an IPO document are written to protect a company from investor lawsuits, so they tend to name every serious danger, including unlikely ones.
Our take
This is not a press release and not a hype video. It is a legal document, filed with a regulator, in which a leading AI lab says that its own estimates point toward AI-driven acceleration of AI “imminently”, and that it has missed such jumps before. That combination is new. It does not mean a self-building superintelligence is around the corner, and Anthropic stresses that no current model has crossed the line. But the debate about AI building AI just moved from conference panels into an SEC filing, and the next question is concrete: what exactly changes in Anthropic’s safety policy when the threshold is reached?
Featured image: Abstract illustration created for ufoks.com, showing a looping AI network and a human hand on the controls.
Sources: Anthropic S-1/A (SEC EDGAR), TechCrunch, Gizmodo, The Latent.



