Anthropic says it may be approaching AI that can improve itself, triggering safety policy review

Anthropic has disclosed in an updated regulatory filing that it may be approaching a critical threshold in AI development: recursive self-improvement (RSI), the point at which an AI system can meaningfully accelerate its own training, research, and capabilities without direct human guidance. The company calls this the most significant safety milestone it has ever defined, and says the threshold may be reached sooner than expected.

The warning appears in Anthropic’s revised S-1A prospectus, filed as the company prepares for a public stock listing targeting November 2026. In the document, Anthropic defines the RSI threshold as the point at which an AI model could perform the equivalent of an entire research team’s work at comparable cost, or at minimum double the current pace of AI progress. No existing Anthropic model has been classified as crossing that line — but the company says the distance to it is shrinking.

Under Anthropic’s Responsible Scaling Policy (RSP), reaching or approaching an RSI-level capability triggers mandatory safety reviews and can require pausing further development until adequate safeguards are in place. The company says it is considering updating the RSP itself to ensure human control is preserved over frontier model development as capabilities grow. That’s an unusual admission: updating the policy that’s supposed to govern development, at the moment when the most consequential threshold in that policy is approaching.

The disclosure also outlines a governance proposal: Anthropic says it plans to convene governments, independent researchers, and rival AI laboratories to test whether coordinated international mechanisms could work in practice to slow or pause development if needed. The company acknowledges that unilateral restraint by one lab accomplishes little if competitors keep building.

The RSI concept has been debated in AI safety circles for years, but this is the first time a frontier lab has publicly disclosed, in a legally binding document, that it believes the capability may be within near-term reach. Critics have noted that the warning serves a dual purpose: it signals Anthropic’s technical seriousness to safety-focused investors while also arguing, implicitly, that the only responsible path is to build the capability inside a safety-focused lab rather than cede the field to others.

Anthropic declined to say which specific capabilities or internal evaluations led to the updated language in the filing, as first reported by TechCrunch.