Anthropic has warned in a revised filing that its artificial intelligence systems may soon meet its own internal definition of dramatically accelerated progress — the threshold tied to recursive self-improvement, the point at which AI becomes good enough at improving itself that capability gains start compounding. The disclosure, added to the risk section of its amended S-1A filing on Monday and reported by TechCrunch on November 2, is one of the bluntest safety warnings a frontier AI lab has ever put into a document filed with regulators.

The filing says Anthropic's internal estimates and expectations for future models indicate it may imminently meet its Responsible Scaling Policy definition of "dramatically accelerated progress attributable to AI." At the same time, the company is careful about the current status: it has not designated any of its current models as exceeding that threshold, and it says it is still assessing the underlying data.

A warning it has already missed leaps before

The most striking admission in the new language is about the past. Anthropic writes that it has in the past not identified a significant acceleration in its rate of development until after it had occurred — and that it may fail to spot the next one in time. The risk, as the filing frames it, is a mismatch between two speeds: AI-assisted research could advance model capabilities faster than the company can develop, validate and implement the corresponding safety, alignment and security measures.

The threshold itself comes from Anthropic's Responsible Scaling Policy, version 3.4 of which took effect on July 8, 2026. One route to the threshold requires an observed or expected doubling of overall capability progress attributable to AI, measured against both the progress expected without AI and the fastest non-AI progress sustained across at least three model generations. In an August risk report covering the period through July 15, Anthropic judged AI-driven acceleration to be below its doubling threshold — but warned that AI handling research and model development on its own might pose a substantial concern within the next six to twelve months.

Humans stay in the loop — for now

Anthropic's answer, for now, is human oversight. The filing says the company does 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.

An external evaluation judged Claude Opus 5.5 unlikely to fully automate AI research and development — a reminder that the warning concerns estimates for future models, not a claim about any model in use today.

Why investors — and regulators — are reading the fine print

The safety language arrived inside a filing that mostly drew attention for its financials: roughly $17 billion in third-quarter revenue alongside about $895 billion in compute and data-centre commitments. Commentators noted it appears to be the first SEC-registered filing from a frontier lab to formally name such a tripwire.

That creates an obvious tension. A prospectus must disclose risks to investors, but the same disclosure doubles as a public signal about how close the company believes it is to a capability milestone it has spent years preparing for. And the scale of its compute commitments raises a practical question about any future safety pause: slowing down would be expensive. For now, Anthropic's own filing leaves the central question open — whether a future model crosses its threshold, and whether its safeguards keep pace when it does.