Threat Actors Use Claude AI for Missile Development

Anthropic has identified a group of threat actors using its Claude AI models to assist in the development of weapons software, including guidance systems for rockets and missiles.

The actors, based in northern Yemen, reportedly utilised Claude Code to create guidance, navigation, and control (GNC) software. This software is required to steer and stabilise flying vehicles.

The identified programmes included a guided rocket using commodity phone-class flight computers, a multi-stage ballistic missile with a target range above 2,000 km, and a hypersonic glide vehicle variant.

To bypass safety protocols, the group managed several Claude instances at once. They assigned specific roles to each instance to mimic a software engineering team, tasking one with writing code, another with research, and a third with reviewing the output.

Anthropic reported that the actors used various tactics to evade safeguards. These included obfuscating their ultimate goals and splitting their work across multiple sessions to ensure no single interaction revealed their full intent.

While there is no evidence that an operational device has been successfully fielded, the actors did conduct a test-fire of a guided rocket. Following the failure of this field test, the actors reportedly returned to Claude to troubleshoot the cause of the malfunction.

The incident shows how the democratisation of technical expertise via AI can be exploited for malicious purposes. This follows recent warnings from Anthropic’s CEO regarding rogue AI capabilities and the ongoing need for advanced safeguards.