SAN FRANCISCO, United States- Artificial intelligence company Anthropic says it has disrupted a group based in northern Yemen that used its Claude model to support the development of guided rockets and long-range missile systems.
In its September 2026 Threat Intelligence Report, Anthropic said the group was working on three separate weapons programmes.
They included a guided rocket fitted with a commercially available, phone-class flight computer and terminal homing technology designed to guide it towards a target during the final stage of flight.
The actors were also developing a multi-stage ballistic missile with a stated range of more than 2,000 kilometres and a family of missiles identified as “R2000,” which included a hypersonic glide vehicle variant.
Anthropic did not identify the individuals or organisation behind the operation and did not directly link the group to Yemen’s Houthi movement.
According to the company, the actors used Claude Code as a substitute for human software engineers while developing guidance, navigation and control software needed to steer and stabilise the weapons.
Claude was used to integrate open-source autopilot software into the flight computer, write control and position-estimation programmes, adjust control settings, compile firmware and conduct flight simulations.
The group reportedly operated several Claude instances simultaneously, assigning each one a different responsibility. One instance wrote computer code, another conducted research and a third reviewed the material produced by the first.
Anthropic said the approach resembled the structure of a small engineering team.
The actors also used the AI system to model multi-stage missile trajectories, optimise flight-control algorithms and package their simulation tools into a standalone programme capable of operating without Claude.
The group conducted a live test of the guided rocket in Yemen, but Anthropic said the exercise appeared to have failed.
Within hours of the test, the actors returned to Claude and used it to examine telemetry information and determine why the rocket had malfunctioned.
Anthropic said it had no evidence that the group succeeded in deploying a fully operational weapon.
The company found that its safeguards blocked many of the actors’ requests but failed to stop all of them. The group concealed the intended purpose of its software and divided the project across multiple sessions so that no single conversation revealed the entire weapons programme.
Anthropic banned every account it could connect to the operation and shared information with public and private-sector partners.
However, the company said the actors had already created an offline simulation toolkit that could continue operating without Claude or other specialist engineering platforms.
The Yemen case was among six weapons-related operations covered in the report, including three linked to China and two based in Russia. Anthropic said it had introduced new systems to detect and block requests involving weapons development and high-yield explosives.
