Anthropic has revealed in a new AI misuse report that a cell of threat actors based in Yemen, understood to be linked to the Iran-backed Houthi movement, attempted to use its Claude chatbot to develop weapons, including ballistic missiles. The disclosure is significant because it shows how a widely used commercial AI model was allegedly exploited to assist with sophisticated military engineering tasks by a US-designated terrorist organisation, despite the company's safety safeguards, raising fresh concerns about the misuse of AI tools in weapons proliferation.

According to the report, the group ran three weapons development programmes, working on a multi-stage ballistic missile, a multi-variant missile and a guided rocket using a commodity phone-based flight computer. Anthropic said the actors used Claude Code to write guidance and control software, coordinating several chatbot instances to divide tasks such as coding, research and review, and evaded safeguards by concealing their goals and splitting work across sessions. While there is no evidence of an operational device being fielded, the group reportedly test-fired a guided rocket that failed, before returning to Claude to diagnose the fault; Anthropic has since blocked the associated accounts, though it warned the actors had already built an offline simulation toolkit independent of its systems. The report also flagged separate misuse in China and Russia involving weapons design software, state-linked surveillance of dissidents, and Russian propaganda generation.