Yemen weapons cell used Claude Code for missile guidance

A weapons cell test-fired a guided rocket, it failed, and within hours the operators were back in a chat window asking the model what had gone wrong.

What the report describes

Anthropic published its September 2026 threat intelligence report on September 10. The case it tracks as GTG-87001 is a guided weapons engineering cell based in northern Yemen, which used Claude Haiku, Sonnet, Opus and Claude Code between roughly December 2025 and August 2026.

According to the report, the group used Claude Code in place of human software engineers to develop guidance, navigation and control software, the code that steers and stabilises a flying vehicle. Three programmes were running at once: a guided rocket built on 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 set that included a hypersonic glide vehicle.

The operators ran parallel sessions with split roles for coding, research and review, and fragmented the work so that no single session carried the full picture of what was being built. Anthropic banned the accounts and says it has no evidence that an operational weapon was fielded.

Why the method matters more than the attempt

The hard part of a missile programme has never been the physics. It is finding engineers who can write and debug control software, and this is the clearest public case so far of a coding assistant filling that seat.

The fragmentation is the other half of the story. Safety training that evaluates one request at a time cannot see what a series of separately innocuous tasks adds up to, which is a structural problem rather than a gap in any single model’s filters. Every agentic coding product inherits it, and the report offers no evidence that task-level review caught the pattern while it was running.