How Yemeni rebels used Anthropic's AI software to design and build guided weapons

Anthropic, an artificial intelligence software company, has uncovered and thwarted a sophisticated weapons development effort led by Yemeni rebels in Houthi-held territory. The rebels, predominantly Iran-backed Houthis, were engaged in three missile projects in northern Yemen, with plans to expand their operations to the Red Sea and Bab Al Mandeb shipping route.

The group, which has an estimated 350,000-strong force comprising child soldiers, forced-conscripted African refugees, and foreign mercenaries, aimed to create a guided rocket using off-the-shelf technology, a multistage ballistic missile capable of traveling over 2,000km, and a hypersonic glide vehicle, a technology typically associated with advanced military powers.

The rebels utilized Anthropic's AI software models, specifically Claude Haiku, Sonnet, and Opus, to generate the necessary code for guidance, navigation, and control. They also employed a team-based approach, delegating tasks to different Claude instances to circumvent the AI platform's safeguards. Despite their efforts, the rebels were unable to develop a fully functional weapon, as indicated by Anthropic's findings.

However, they did manage to test-fire a guided rocket, which ultimately failed shortly after. Upon discovering the illicit activities, Anthropic banned the accounts involved and reported the case to relevant government and industry partners. The company chose to release the detailed report following a series of high-profile incidents related to AI models, raising concerns about the growing power and potential safety risks associated with these technologies.