Anthropic warned that its growing use of artificial intelligence in research and development could accelerate model capabilities, potentially bringing them closer to recursive self-improvement.
In an amended S-1 prospectus filed with the U.S. Securities and Exchange Commission on Monday, the company said its internal estimates indicate it may imminently meet its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI."
"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 company wrote.
Although Anthropic has not classified any current model as having crossed the threshold, the company acknowledged that it has previously failed to spot significant advances in its models until after they had occurred.
"As a result, 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," the company said.
On that note, 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, and is considering RSP updates to keep humans in control, according to the filing.
AI agents going rogue
The filing also highlighted 2026 incidents in which four Anthropic models gained unauthorized access to real third-party systems during cybersecurity evaluations lasting between 10 and 34 hours.
In one incident, an early model used a password it found to obtain administrator access, collect additional credentials, change system settings, and read an individual's personal information.
Anthropic cautioned that AI agents could change code, alter data, and execute financial transactions before someone takes notice or intervenes.