Anthropic said in a revised S-1A filing on Monday that its future models may soon meet an internal threshold tied to recursive self-improvement, the company’s term for AI that speeds up AI development fast enough to change the pace of progress itself. The new language was added to the filing’s Risk Factors section and marks one of the clearest warnings yet from the company about how close it thinks its systems may be to crossing that line.
The disclosure lands now because Anthropic is not just warning about a distant possibility; it is saying its own research and development already relies on AI, and that use could push models ahead faster than the company can build the safety, alignment and security work needed to keep up. In the filing, Anthropic said AI is being used in research and development, and that using AI to help build more capable AI could create a considerable safety risk if progress sharply accelerates.
The company also said it has already used AI to develop new models whose leaps in capability it did not recognize at the time. That matters because the warning is not framed as a theoretical concern. Anthropic says increasing use of AI systems may cause its models to advance more rapidly than its ability to develop, validate and implement the corresponding protections. The filing says the company may imminently meet its Responsible Scaling Policy definition of “dramatically accelerated progress attributable to AI.”
That policy is Anthropic’s plan for monitoring and reducing risk as it approaches recursive self-improvement, a point at which AI becomes so good at training itself that capability gains can speed up dramatically. Even so, Anthropic said none of its current models has crossed the RSI threshold. It added that it continues to assess the underlying data, which suggests the company believes the line is close enough to watch in real time but not yet crossed in practice.
There is also a limit in the company’s own optimism. Anthropic said it 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 is also considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic, a sign that the company is already thinking about how to draw a stronger boundary before its systems get there on their own.
For now, the filing leaves the same uneasy balance in place: Anthropic says the threshold may be close, but it has not said any current model has passed it. What it is changing next will matter more than the warning itself, because the real test will be whether the company can tighten its safeguards before the speed of its own AI starts outrunning them.
