AI use in guided weapons software: GTG-87001 case

What happened

We identified a threat actor group based in northern Yemen that ran three weapons development programs: a guided rocket with terminal homing guidance using commodity phone-grade flight computers; a multistage ballistic missile with a claimed range target of over 2,000 km; and a multi-variant missile (called the 'R2000' series) that includes a hypersonic glide vehicle variant.

What the AI did

The actors used Claude Code in place of human software engineers to develop guidance, navigation, and control (GNC) software used to guide and stabilize aerial vehicles. For example, they used Claude to integrate an open-source autopilot onto a phone-grade flight computer, write control and position estimation software, tune control settings, run firmware build pipelines, and execute flight simulations. The actors managed multiple Claude instances simultaneously, assigning each a role as a manager would delegate work in a small engineering team: the actor assigned one instance to write code, another to conduct research, and a third to review the code produced by the first.

Our safeguards blocked many of their requests, but not all. The actors used various strategies to evade our safeguards, including hiding the intended purpose of their targets and software, and spreading work across multiple sessions so that no single session revealed their full intent.

What the report observed

These actors continuously worked to develop guided weapons, including using Claude to design guidance software. We have no evidence that the actors successfully deployed an operational device; however, they did test-launch a guided rocket. This field test appears to have failed: within hours, the actors returned to Claude to analyze the cause of the failure.

We identified the activity as part of an internal investigation into suspected weapons development. We banned accounts associated with the actors and shared threat information with public and private sector partners to mitigate the risk posed by the actors. However, we have evidence that the actors have built an offline simulation toolkit that does not rely on Claude or other engineering computing environments such as MATLAB.

This site does not publish any weapons software code, control parameters, or information that could be used to manufacture weapons.

Confirmed & unknown

Platform response

Anthropic banned accounts associated with the actors and shared threat information with public and private sector partners.

Limits of response：Banning accounts cannot stop the actors from continuing weapons development using the offline toolkit.