The Houthis and AI-assisted ballistic proliferation

‘An engineering cell based in Yemen was using Claude to develop software for guided weapons,’ announced Anthropic (the company developing the large language models (LLMs), Claude) in a security report entitled ‘Detecting and countering misuse of AI: September 2026’.1 The unit, identified by the company under the code GTG-87001, used Claude to design three different programmes:

• Final-phase homing guidance software for a guided rocket intended to operate on a commercial off-the-shelf smartphone.
• A multi-stage ballistic missile with a target range of at least 2,000 km.
• A multi-variant missile, one variant of which would incorporate a Hypersonic Glide Vehicle (HGV).

Claude’s operating mode was particularly well-oiled: queries were broken down as much as possible and functions separated so that safeguards could not identify the purposes associated with them and would not censor them. For example, and again according to Anthropic, some accounts carried out the searches, others asked Claude to write code, and finally, others corrected the code. This team employed what is known as ‘fragmentation evasion’. After linking all the prompts together, the company had these actors’ accounts closed, whilst admitting that it had evidence that they already possessed an offline simulation capability – unrelated to Claude – enabling them to continue their numerical simulations. Furthermore, although the accounts have been closed, the Yemeni actors will still be able to create new ones and, once again, attempt to circumvent the safeguards. Even if a direct link to the Houthis has not been established, it nevertheless seems reasonable to attribute this body of research to them, given that they control the region from which the research was launched and have been actively pursuing a strategy of ballistic weapons proliferation for over a decade.

The question, therefore, is whether the use of an LLM by a Houthi engineering unit simply constitutes a new technical support tool, or whether it reveals a more profound transformation in the capability-building model of an actor subject to an embargo?