How Yemeni Rebels Leveraged Anthropic AI to Design Guided Weapons

Anthropic revealed in a comprehensive report this week that it successfully investigated and disrupted threat actors utilizing its AI software for military objectives. One such operation was identified in Yemen, specifically within Houthi-held territory. Titled "Detecting and countering misuse of AI", the report details an ambitious weapons development initiative in northern Yemen, where a faction pursued three sophisticated missile projects simultaneously. Iran-backed Houthis maintain control over the capital Sanaa and northern Yemen, which has increasingly served as a staging ground for attacks against neighbours. Recently, they captured the strategic port city of Mokha on the western coast, broadening their dominance along the Red Sea and close to the Bab Al Mandeb shipping route.

The findings offered extraordinary insight into the scale and complexity of the group's military ambitions. One project involved converting commercially available technology into a guided rocket, utilizing an off-the-shelf, phone-grade flight computer alongside a homing system intended to direct it toward a target during its final flight phase. A second, far more ambitious undertaking focused on a multi-stage ballistic missile designed to travel over 2,000 kilometres. The most notable effort, however, was designated the "R2000" programme, encompassing proposed missile variants that featured a hypersonic glide vehicle—a technology typically associated with the world's premier military powers.

"The activity took place between December 2025 and August 2026, involved Claude Haiku, Sonnet, and Opus models, and was disrupted by Anthropic, which banned the accounts and published the case in the report," the company stated. According to Anthropic, rather than employing human software engineers, the actors utilized Claude Code to author the guidance, navigation, and control software necessary to stabilize vehicles and maintain their course. They structured their workflow cooperatively, running multiple Claude instances simultaneously with distinct responsibilities: one generated code, another conducted research, and a third evaluated the output, while human operators oversaw the process and distributed tasks.

Although Anthropic's safeguards rejected numerous prompts, the group bypassed certain restrictions by masking their objectives and splitting tasks across separate sessions to prevent any single conversation from revealing the full scope of their work. While they did not produce a fully operational weapon, they managed to test-fire a guided rocket. The test appeared unsuccessful, as the operators returned to Claude shortly afterward to diagnose the failure. Anthropic uncovered the activity via internal investigations into suspected weapons development, subsequently banning the accounts and sharing threat data with government and industry partners. Nevertheless, because the group had already constructed an offline simulation toolkit operating independently of Claude or tools like MATLAB, a portion of their capability persisted past the bans.

The release of Anthropic's report follows mounting concerns regarding the safety risks and escalating power of advanced AI models. Over the past decade, the Houthis have evolved from a localized armed faction into a formidable regional threat, targeting shipping lanes and neighbouring states while solidifying their hold over Sanaa and north-western Yemen. Estimates from the US-based Middle East Institute place Houthi troop strength at 350,000, including child soldiers, forcibly conscripted African refugees, and foreign mercenaries. The rebels depend heavily on cross-border smuggling networks and intricate maritime supply routes traversing the Arabian Sea, the Gulf of Aden, and the Red Sea to sustain their logistical and military operations, acquiring weapons, drone components, ballistic missiles, and naval technology.