Anthropic warns that it’s very close to RSI capability, aka self-building AI

Anthropic filed a revised S-1A on Monday that drew attention for its updated financials — revenue climbing sharply, but expenses climbing too. Buried deeper in the filing, though, was a second bombshell: newly expanded risk-factor language suggesting the company may be approaching recursive self-improvement, or RSI.

RSI describes the point at which an AI becomes capable enough at training itself to produce a dramatic acceleration in AI capabilities. For critics of the technology, that threshold marks a potential point of no return, where a system improves beyond human control. For proponents, it marks the arrival of a productivity and discovery engine of enormous scale. Anthropic's filing treats the possibility as both promise and peril.

The company acknowledged in the S-1 that it already uses AI to accelerate its own research and development, while cautioning that doing so carries real risk: "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 ties that risk to Anthropic's Responsible Scaling Policy, which sets out how it plans to monitor and mitigate harms as RSI approaches.

Two specific disclosures in the Risk Factors section stand out. First, Anthropic said it has already used AI to develop new models whose capability jumps it did not recognize at the time — and warned it may miss future accelerations until after they occur. Second, citing internal estimates and expectations for future models rather than current ones, the company said it may imminently meet its own RSP definition of "dramatically accelerated progress attributable to AI." It has not designated any current model as exceeding that threshold and says it continues to assess the underlying data.

Anthropic also committed to keeping people in the loop. It said it does not anticipate safely using Claude to design and build successive frontier model generations fully autonomously without meaningful human supervision, "if and when that becomes technically possible," and is considering RSP updates to keep humans in control of frontier development.

The open questions now are concrete. Whether Anthropic's internal data confirms the threshold crossing, how any RSP amendment changes its disclosure and deployment cadence, and whether other frontier labs revise their own risk language in response will all signal how close the industry actually is to self-improving systems.