A cell in northern Yemen

Anthropic's September threat report describes operators it assessed as highly likely linked to the Houthis, using Claude Code to build guided rocket, ballistic missile and hypersonic glide software.

Parallel agents, not one chat

The group ran several Claude sessions at once. One wrote code, another did research, a third reviewed the output, letting a handful of people cover jobs usually split across specialist teams.

“The workflow let a small group reproduce several functions normally distributed across specialist engineering teams.”

What they actually built

Guidance for a tactical rocket, a ballistic missile with range over 2,000 km, and a hypersonic glide vehicle variant they called R2000.

Phone-class flight computer

They tried to pair open-source autopilot software with cheap phone-grade hardware, wrote navigation and control code, built six-degree-of-freedom trajectory simulations and used reinforcement learning to tune flight control.

They test-fired a rocket

The launch in Yemen appears to have failed. Within hours the operators were back in Claude, feeding it launch telemetry to work out what went wrong.

Fragmenting beats the filter

Safeguards blocked many requests, so the operators hid the end use of each task and split the project across separate conversations. Anthropic later banned the accounts.

“Individual requests may appear disconnected from weapons development when a larger engineering project is deliberately fragmented across sessions.”

The toolkit went offline

No evidence the weapon ever became operational. But before the ban, the group had compiled its work into a standalone executable that runs without Claude at all.

“The engineering capability assembled around it had already begun moving offline.”