Anthropic report says Yemen-based armed group used Claude for missile guidance software
The company's September report describes three weapons programmes, a guided missile test that reportedly failed, and accounts that have since been banned.
Three weapons programmes
Anthropic has acknowledged that actors linked to an armed group in northern Yemen used its Claude model to write guidance software for missiles and ballistic missiles. The findings come from a report the company published in September 2026. According to the report, the group ran three weapons programmes. One was a guided missile whose flight computer was of the kind found in ordinary phones and which steered itself during the final phase. A second was a multistage ballistic missile with a targeted range of over 2,000 kilometres. The third was a missile set called "R2000" that included a hypersonic glide vehicle variant.
The Financial Times reported that the users could be members of the Iran-backed Houthi militia, which controls large parts of northern Yemen, according to Tages Anzeiger. Digi24 reports that Anthropic linked the actors to the Houthi group, while Tages Anzeiger notes that the company does not name the Houthis explicitly. The Houthis have long fought Yemen's internationally recognised government and have also attacked Saudi Arabia, aiming to block its oil exports.
Parallel instances of Claude
Instead of human software engineers, the group reportedly ran several instances of Claude at once, assigning each one a role. One wrote code, a second carried out research and a third checked the code the first produced. Anthropic said the actors had Claude integrate an open-source autopilot into a phone-style flight computer, write control and position-estimation software, tune control settings, run a firmware compilation pipeline and carry out a flight simulation. The Daily Mail, as cited by wnp.pl, reports that the model also adjusted control parameters and made changes when missiles failed to hit the target. Anthropic's report reportedly said the AI condensed the research of leading experts on missile technology, which the actors then accessed.
Safeguards and split sessions
Anthropic said its safeguards blocked many of the group's requests, but not all of them. The actors hid their objectives and the applications the software was meant for, and split the work across several sessions so that no single session revealed their full intent. Wnp.pl, citing the Daily Mail, reports that the company's authors described a range of tactics the group used to get around the protections. Wnp.pl also says Anthropic only realised what its software was being used for after the work was underway, and that the group then used the knowledge it had gained to make its own changes. Wpolityce reports that the group eventually built offline simulation tools, which gave it independence from external servers.
Failed test and banned accounts
Tages Anzeiger, citing The National, reports that at least one of the guided missiles was actually tested and that the test apparently failed. A few hours later the actors turned to Claude again to find out what had caused the failure. Anthropic then blocked the affected accounts and passed information about the activity on to competent authorities and partners. Whether the group was able to develop missiles fit for use is not known.
- Anthropic publishes its misuse report describing the Yemen-based group's activity
- A guided missile test reportedly fails, and the actors return to Claude hours later
- Anthropic blocks the affected accounts and passes information to authorities and partners
Riyadh attacks and open questions
Wnp.pl, citing the Daily Mail, says it is unclear whether the group used Claude in weekend attacks on an airport in Riyadh in early October. Those attacks reportedly killed 14 people and wounded more than 300, although a wpolityce headline gives the death toll as 12. According to the Daily Mail as cited by wnp.pl, the strikes shocked Saudis with their range and precision and prompted discussion in the White House about expanding US military support for Saudi Arabia. Tages Anzeiger notes that experts see it as especially problematic that AI systems can now take over technical tasks that previously required specialised experts.
