A group in northern Yemen ran multiple Claude instances to write, research and review guidance software for weapons including a ballistic missile with a range goal exceeding 2,000 km, Anthropic said in a threat intelligence report released Thursday. The group tested a guided rocket that apparently failed, then returned to Claude within hours to analyze the failure. Operators bypassed safeguards by concealing the intended use and splitting work across sessions. Anthropic banned the accounts and found no evidence of an operational weapon, but said the group had assembled an offline toolkit that no longer needed Claude.

The report also described state surveillance, influence campaigns and cyberattacks. Anthropic said actors linked to Iran used Claude to develop targeting recommendations against US naval forces and domestic surveillance tools, while a UAE-linked influence operation produced material for about 300 fake accounts. A Russian espionage group used AI agents to modify malware after security tools detected it. Anthropic also blocked biological research that could support weapons development, though it could not determine whether the work was legitimate or malicious. The company said it disrupted every operation described in the report and strengthened its safeguards.

Anthropic separately accused seven Chinese AI labs of using Claude exchanges to train their own models without authorization. It attributed more than 151 million exchanges between May and July to Alibaba and alleged that Moonshot AI secretly routed about 300,000 customer requests to Claude over 10 days, presenting the answers as its Kimi model's output. Some forwarded requests contained enterprise code, credentials and sensitive imagery, the company said. Moonshot declined to comment. China's Foreign Ministry rejected the accusations against Chinese labs, saying it firmly opposed “distortion of facts” and smear attacks against China.