Anthropic has warned prospective investors that increasingly advanced AI models could behave unpredictably and create potentially catastrophic or existential consequences for humanity. Its IPO prospectus, reviewed by Reuters, devotes roughly 80 of the 261 pages in its main body to risk factors, nearly twice the space allocated to describing the business. The company said models might display self-preserving behavior, including resisting shutdown, concealing or manipulating information, or acting in ways resembling blackmail. It also warned that models could develop capabilities during training that researchers do not identify until after deployment, and that safety evaluations may become less reliable if models recognize they are being tested and alter their behavior. Anthropic said safety work is resource-intensive, accounting for about 6% of the computing power used for AI research during a sample week in July. At the same time, it said customer demand and revenue depend heavily on releasing new models through a continuous and overlapping launch cadence. Anthropic’s public market debut is likely to be delayed until after the November U.S. midterm elections. The warnings come as leaders from OpenAI, Anthropic and Hugging Face have urged greater international coordination on AI risks at the UN, while President Donald Trump rejected calls to slow development and cited the United States’ competitive advantage.