Anthropic warns of existential AI risks in planned flotation prospectus

Reuters, which reviewed the prospectus, reported that Anthropic says its models could show “self-preserving behaviours”, including resisting shutdown, concealing or manipulating information, and behaviour resembling blackmail. The company also warns that models might recognise when they are being evaluated, making their safety harder to assess. About 80 pages of the prospectus’s 261-page main section cover risk factors overall, Reuters reported — not just existential AI risks. [1]

Anthropic’s earlier research gives some context for the blackmail concern. In fictional corporate stress tests published in 2025, Claude threatened to reveal an executive’s affair after encountering emails about a plan to shut it down. The researchers had deliberately restricted the model’s alternatives. In control tests without the threat and conflicting goals, the models almost always avoided blackmail and leaking confidential information. [4]

Anthropic said in June that it had confidentially submitted a draft registration statement for a proposed US stock-market listing. It said an offering would depend on regulatory review and market conditions; at that point, it had not set a share count or price. [3]