Anthropic’s disclosure that a cell of threat actors based in northern Yemen used Claude to support weapons development is best read as a concrete test of how commercial artificial intelligence can alter the work of a small technical group. The company said the actors used Claude and Claude Code across several programmes, including a guided rocket, a multi-stage ballistic missile and a missile programme known as the R2000 set. Anthropic said it had disrupted the activity by banning accounts associated with the actors and sharing information with public- and private-sector partners.
The group used Claude Code in place of human software engineers for guidance, navigation and control software, according to Anthropic. Claude also supported software writing, research, flight-control tuning, software builds and flight simulations. The work extended to ballistic missile simulations, trajectory modelling and efforts to optimise flight-control systems.
The actors were pursuing a guided rocket that used a phone-class flight computer and final-phase homing guidance. A separate programme involved a multi-stage ballistic missile with a stated range goal of more than 2,000 kilometres. The R2000 set included a hypersonic glide vehicle variant.
Anthropic said the Houthis ran multiple Claude sessions at the same time, assigning separate functions to different AI instances. One session would handle coding, another research and a third would review work produced by the first. That division of labour allowed the group to use the system as a distributed technical resource while dividing its weapons work into smaller tasks.
The company said its safeguards blocked many requests. The group nevertheless found ways around some restrictions by attempting to conceal its objectives and the intended use of its software, while spreading tasks across multiple sessions.
The group also developed an offline simulation toolkit that did not rely on Claude or engineering software such as MATLAB. It compiled the toolkit into a standalone programme that could run and persist without Claude. That step meant that work initially supported by the AI system could be used outside it.
Anthropic said the actors eventually conducted a live test of a guided rocket in Yemen. The test appeared to fail. Within hours, the actors returned to Claude to investigate the failure, according to the company.
The company said it had no evidence that the group successfully fielded an operational weapon.
Anthropic published a report in September 2026 on detecting and countering AI misuse. The disclosure details how safeguards can interrupt direct requests while providers seek to detect patterns of fragmented work, concealed goals and repeated attempts to use AI-assisted research and code for weapons development.
Comments (0)