Anthropic flags AI R&D acceleration risk in amended IPO filing

Anthropic added new risk language about recursive self-improvement and AI-driven research-and-development acceleration to its amended S-1/A registration statement filed with the U.S. Securities and Exchange Commission for its IPO, according to TechCrunch. The event date of the filing is November 2, 2026.

In the filing, Anthropic said model capabilities could improve faster than the company can develop and verify its safety, alignment and security measures. It also acknowledged that in the past it only recognized a major acceleration in development speed after the change had already occurred, and said this could happen again.

Anthropic said its internal estimates and outlooks for future models suggest it may soon reach the threshold for dramatic AI-driven development acceleration, while no model has yet been judged to cross it. Under its Responsible Scaling Policy, that threshold is defined as overall capability progress more than twice as fast as prior expectations and historical bests, largely due to automated AI R&D.

The company said it is reviewing revisions to its safety policy to keep meaningful human oversight and control over key decisions in developing successor models. It noted that Claude has been used for software writing, optimizing model training code and research experiments.

Earlier disclosures

Anthropic's initial filing, on October 26, 2026, already mentioned the possibility and risks of RSI. In September 2026, the company stated in its filing that Claude wrote more than 90% of the code contributed to its internal codebase.