Yemen group used Claude AI to build guided rocket: Report

The Yemen operation was one of six cases in which Anthropic said its models had been used in weapons-related work.

Hyderabad: An engineering group in northern Yemen used Anthropic’s Claude AI models to write the guidance software for a rocket it later test-fired, the company said in a report published on Thursday, September 10.

The test appeared to fail. Within hours, the operators were back on Claude trying to work out what had gone wrong.

The Yemen operation was one of six cases in which Anthropic said its models had been used in weapons-related work. The company also identified separate cases involving surveillance systems developed for government agencies.

One tap — then press “Add” on Google

The report covers activity Anthropic detected and shut down between December 2025 and August 2026. The company said it banned all the accounts involved and shared information with government and industry partners.

Three weapons programmes

Anthropic said the Yemen-based group was working on three weapons programmes at the same time. One involved a rocket built around a phone-grade flight computer. Claude was used to write the software responsible for steering and stabilising the vehicle during flight.

The group was also developing a multi-stage ballistic missile with a stated range of more than 2,000 km, as well as a missile system that included a hypersonic glide variant. The operators split the work between several instances of the model, giving each a different job. 

One was used to write code, another to conduct research and a third to review the work.

Anthropic said it found no evidence that the group had deployed a working weapon. But some of the work had already moved beyond the chatbot itself, which means that the group had compiled its simulation tools into a standalone programme that could continue to run without Claude.

Drone swarm and Taiwan targets

The company described two other weapons-related cases in similar detail.

In one, Russia-based developers used Claude to build software for a swarm of first-person-view kamikaze drones. The system was designed to allow an onboard model to select targets, including people, and trigger a detonation without a human making the final decision.

Anthropic said the team trained the system on combat footage scraped from Ukraine and repeatedly used a fixed location in Donetsk Oblast as a demonstration strike point. 

It assessed the developers as a small freelance group rather than a Russian state organisation.

In another case, a China-based researcher used Claude to build a suite of targeting tools for electronic warfare and attacks on enemy air-defence systems. During the work, Anthropic said, the user changed the default simulation to 12 targets in Taiwan. They included a command bunker, an early-warning radar site and Patriot missile batteries.

Account information linked the user to Chinese research institutions, including the People’s Liberation Army Academy of Military Sciences, according to the report.

Mali-based phone monitoring system

Anthropic’s surveillance findings centred on a system it called Lakana 360, which it said had been built for Mali’s state intelligence agency by what appeared to be a single consultant in Bamako.

The platform covers all three of Mali’s mobile networks and about 25 million SIM cards. It collects call records, text messages and voice traffic, can identify people by their voices across different SIM cards, flag users of encryption tools and match individuals against the country’s biometric register.

Anthropic said a requirement for a court order was removed, at the operator’s request, from the part of the system that generates intelligence reports on individual phone numbers.

The company also pointed to reporting by the US State Department and Human Rights Watch documenting the detention and abduction of opposition figures, journalists and civil society members by Malian security forces.

Banning the consultant’s account did not shut down the surveillance platform. Anthropic said it was running on local servers using other AI models, meaning the company’s enforcement action did not affect the system already in operation.

Across the cases, Anthropic said its safeguards blocked many requests but did not stop all of them. Users were able to get around some restrictions by concealing what they were building and breaking the work into separate sessions.