Scientists and state actors evaded safeguards to use Claude for biological weapons research and mass surveillance, Anthropic said in a report released Thursday. The AI company blocked researchers whose work appeared linked to military research and banned accounts tied to state-run surveillance programs globally. Anthropic also accused China's Moonshot AI of covertly routing thousands of user requests to Claude models and presenting the responses as its own Kimi model.

The findings include the use of AI to write guidance software for a 2,000 km range ballistic missile in Yemen and the design of an autonomous kamikaze drone swarm by a Russia-based team. The report also identified a campaign linked to Alibaba that performed 151 million exchanges for illicit model distillation. Anthropic shared the results with authorities and other AI companies to strengthen industry safeguards.