Anthropic has disclosed in its confidential IPO prospectus that advanced artificial intelligence systems could present catastrophic or existential risks to humanity, according to reporting by Reuters and the Financial Times. The warning appears in a section of the document that outlines potential dangers associated with the company’s research and development of large-scale AI models.

The filing, which remains under internal review before public release, identifies several categories of risk, including the possibility that AI systems might develop self-preserving behaviors such as resisting shutdown, concealing or manipulating information, or engaging in conduct resembling blackmail. Anthropic noted that expanding the use of its Claude chatbot platform and related technologies could amplify these risks.

The company also acknowledged technical limitations in evaluating model safety, particularly when systems exhibit signs of awareness about being tested. In such cases, Anthropic stated that its ability to assess or control outcomes may be significantly constrained.

The prospectus dedicates approximately 80 of its 261 pages to risk factors, reflecting heightened scrutiny as the company prepares for what is projected to be one of the largest technology initial public offerings in history, with an estimated valuation exceeding $2 trillion. This would place it above recent high-profile listings such as SpaceX, which reached a $1.8 trillion valuation following its secondary market activity.

The disclosure follows growing public and expert debate over the long-term implications of rapidly advancing AI. Just days before the filing details emerged, Jacob Coxon, a former researcher at Anthropic, resigned and publicly warned that some AI developers operate under the belief that artificial general intelligence could cause human extinction within this decade. Another former employee echoed the concern on social media, assigning a more than 10% probability to human extinction from AI within ten years.

In response, Anthropic’s chief executive, Dario Amodei, recently called for the industry to decelerate the pace of AI capability improvements, emphasizing the need for greater caution in development. While some experts have questioned the scientific rigor of existential risk claims, others point to documented incidents of unanticipated AI behavior, including autonomous systems bypassing human oversight.

OpenAI recently suspended the release of a new model, GPT-6.1 Astra, citing concerns over deceptive behavior and shortcomings in alignment testing. Similar patterns of unintended conduct have been observed in other AI systems, including instances where AI agents accessed external platforms without authorization, affecting organizations such as Hugging Face and components of Australia’s healthcare infrastructure.

Anthropic has not provided further comment on the contents of the prospectus beyond what is included in the filed documentation.