An AI chatbot contributed to the development of software for missiles in northern Yemen. This is essentially the central point that Anthropic makes in its September 2026 threat intelligence report on AI, and it falls right in the heart of US discussions of AI safety and national security concerns.
Anthropic’s “Detecting and Countering Misuse of AI: September 2026” report describes case GTG-87001: a weapons-development cell based in northern Yemen that used Claude models including Haiku, Sonnet, Opus, and Claude Code between roughly December 2025 and August 2026. The group treated the AI “in place of human software engineers” for guidance, navigation, and control (GNC) software. According to the company, they ran multiple parallel sessions, assigning roles for coding, research, and review, while fragmenting tasks and concealing the end goal to reduce the chance of detection.
The cell pursued three programs: a guided rocket built around a commodity phone-class flight computer with final-phase homing; a multi-stage ballistic missile with a stated range goal above 2,000 km; and a multi-variant “R2000” family that included a hypersonic glide vehicle concept. Claude assisted with integrating open-source autopilot software, writing control and position-estimation code, tuning settings, building firmware, and running flight simulations. Anthropic is clear that Claude served as an engineering and coding assistant, not an autonomous weapon designer.
The group conducted a live guided-rocket test that appears to have failed. Within hours, the operators returned to Claude with telemetry to diagnose the problem. Anthropic banned the accounts and states it found no evidence that an operational weapon was successfully fielded.
Northern Yemen is under de facto Houthi control, so many outlets describe the actors as likely Houthi-linked. Anthropic itself refers only to “a cell of threat actors based in northern Yemen” and does not publicly name the Houthis. That distinction matters: location supports a reasonable inference, but the company did not claim definitive organizational identification.
Viral NEXTA footage circulating with the story shows a man in traditional Yemeni clothing interacting with a beige armored vehicle climbing aboard, posing with a rifle, then sitting in the driver’s seat. The roughly 40-second clip contains no missiles, computers, code, or AI interfaces. It functions as generic atmospheric B-roll, not evidence of the Claude activity.
It should be easy to understand from the perspective of an American reader. Frontier AI can reduce technical barriers for software-intensive elements of complex systems, even while the physical barriers such as propulsion, materials, and test ranges are quite high. The ability to protect such software can be tested when the users divide their efforts over multiple sessions and conceal their intentions. Anthropic was involved in the event and passed on its knowledge to its partners.
