Anthropic said it blocked accounts in Houthi-held northern Yemen after finding they used Claude to help develop a guided rocket and other weapons concepts. The company said it shared its findings with partners, and that the users did not manage to field an operational device.
The key detail is the workflow: after a failed guided-rocket test, the same users went back to Claude to ask why it failed and what to change next. That turns the chatbot into a rapid feedback tool for weapons prototyping, not just a place to generate text or code.
For operators of hosted LLMs that help with coding, troubleshooting, or design, the exposure is the iterative loop itself. If users can turn the model into a technical lab partner, harmful work can speed up even when the first attempt does not work.