A group in northern Yemen, thought to be the Houthis, who this week captured Mayun Island, have used Claude to help write software that guides and steers weapons.
Instead of employing a human engineering team, Anthropic's new report, 'Detecting and Countering Misuse of AI: September 2026', says the group used several Claude Code sessions, with one writing weapon steering software, another researching technical questions, and a third checking the others' work. The software was meant to keep the weapons stable, determine their position, and guide them during flight.
Anthropic's safety systems rejected many of the requests, but some worked because the group concealed their military purpose and divided the project among separate conversations.
The group test-fired at least one guided rocket, but the test seems to have failed. There are also concerns that the group had already created its own offline simulation software, so it may be able to continue some development without Claude.
When compared to Anthropic's previous threat report, there are now more people/groups that are using AI for various types of harm, and it is clear that AI safety has become increasingly tied to geopolitical security.
Jacob Coxon's argument on resigning from Anthropic this week - that frontier labs are expected to strengthen controls while racing to release more capable systems faster than their competitors - applies here as well. This conflict could be helping groups like the Houthis to avoid the safeguards in place and build their weapons programmes.
Voluntary commitments won't stop the Houthis from using AI, which is why we need something closer to a binding international framework.