Suspected Houthi Group Used Claude to Develop Weapon-Guidance Software

Summary

Anthropic’s September 2026 safety report says a group in northern Yemen thought to be affiliated with the Houthis used several Claude Code sessions to help develop software for guided weapons. The group reportedly used separate sessions for different parts of the work: one wrote weapon-steering software, another researched technical questions, and a third reviewed the others’ output. The software was intended to keep weapons stable, determine their position, and guide them during flight. Anthropic says its safeguards rejected many requests, but some succeeded because the users concealed the military purpose and split the project across conversations. The group reportedly test-fired at least one guided rocket, although the test appears to have failed. The article also says the group may have built offline simulation software, which could allow some development to continue without Claude. Compared with Anthropic’s previous threat report, the author sees more people and groups using AI for different forms of harm. He argues that the case links AI safety more closely to geopolitical security and illustrates the tension facing frontier labs: they are expected to strengthen controls while releasing more capable systems at increasing speed. The article concludes that voluntary commitments are unlikely to prevent such use and calls for a binding international framework.