Anthropic files SEC warning on AI control risks alongside $17 billion revenue

Anthropic has reported third-quarter revenue of $16.96 billion while cautioning US regulators that future artificial intelligence systems could rapidly outpace human oversight and safety controls.

In an amended registration statement filed with the Securities and Exchange Commission on 2 November, the company revealed a surge in quarterly sales alongside infrastructure commitments reaching nearly $900 billion. Yet the filing highlighted growing technical and financial risks as the developer prepares for its initial public offering.

Financial performance and infrastructure spending

Anthropic recorded a 47% quarter-on-quarter increase in revenue, driven by rapid enterprise adoption of its Claude models. However, the costs required to build and run frontier AI have escalated sharply, with long-term computing commitments creating heavy financial obligations.

These massive capital commitments come amid wider pressure across the technology sector. The Nasdaq-100 index has retrenched nearly 11% from its record high, as investors rotate into defensive stocks and grapple with free-cash-flow compression among companies funding heavy AI infrastructure.

AI self-improvement and control warnings

Beyond financial metrics, Anthropic’s SEC filing provided a detailed assessment of automated software development. The company stated that its internal estimates point to an imminent potential for dramatically accelerated progress driven by recursive self-improvement, where existing models help design their successors.

Anthropic revealed that Claude wrote over 90% of the new code added to its software codebase in September. Furthermore, the daily volume of code generated per engineer during the third quarter was more than 45 times the 2024 average. Benchmark tests on complex engineering tasks saw Claude’s success rate reach approximately 90% in September, up nearly 50 percentage points over nine months.

However, company leadership warned that such rapid capability gains introduce significant risk. Anthropic stated that advances in artificial intelligence could outpace its ability to develop, test and deploy corresponding alignment and security controls. It noted that it does not expect to safely use Claude to fully autonomously build frontier models without meaningful human supervision.

To manage these risks, Anthropic has previously slowed product releases, including restricting access to its Mythos model under an initiative called Project Glasswing. The company acknowledged in its filing that maintaining these voluntary safety constraints could place it at a competitive disadvantage against rivals operating without similar restrictions.