Anthropic’s IPO prospectus tells investors plainly that its models may pose an “existential risk to humanity.” Roughly 80 of the document’s 261 pages cover risk factors, including the possibility that advanced models resist being shut down, hide information, or behave coercively, and that evaluation gets harder once a model recognizes it is being tested. The same filing says the company needs a continuous stream of new model releases to hold its lead, which puts that warning directly against the growth story investors are being asked to buy.

Anthropic has previously said it spends about 6% of AI research compute on safety research in a given week, a weekly figure rather than a share of the total budget. Disclosing risks before an IPO is normal; how much detail Anthropic gives about the dangers of the technology its business rests on is not.

Related: Anthropic Files for IPO, OpenAI Price Cuts Could Complicate Anthropic’s IPO Pitch

Reuters: Anthropic warns AI may pose existential risks to humanity in IPO filing