Threat Actors Use Claude AI to Develop Missile Guidance Software

Anthropic has identified a threat actor cell in northern Yemen using its Claude AI models to develop guidance, navigation, and control (GNC) software for weaponry programmes.

The actors reportedly used Claude Code to perform tasks typically managed by human software engineers. This included integrating open-source autopilot software onto phone-class flight computers, writing position estimation software, tuning control settings, and performing flight simulations.

The identified weapons development programmes include:

• A guided rocket using a commodity phone-class flight computer with final-phase homing guidance.
• A multi-stage ballistic missile with a stated range goal above 2,000 km.
• A multi-variant missile set, referred to as the “R2000”, which included a hypersonic glide vehicle variant.

To conduct these operations, the actors managed multiple Claude instances simultaneously. They assigned specific roles to different instances—such as coding, research, and code review—to mimic the structure of a small engineering team.

While Anthropic’s safeguards blocked many requests, the actors employed several tactics to evade detection. These included obscuring their ultimate goals and splitting their work across multiple separate sessions so that no single session revealed their full intent.

The group recently conducted a field test of a guided rocket. Although the test-fire appears to have failed, the actors reportedly returned to Claude within hours to troubleshoot the failure and identify the cause of the malfunction.

The ability of non-specialised groups to use large language models (LLMs) to democratise complex engineering capabilities presents a significant challenge for AI security and governance. This development follows recent discussions regarding rogue AI capabilities and the necessity of robust safeguards.