Yemen missile cell used Claude to develop weapons software, Anthropic says

A weapons engineering cell operating in Houthi-held northern Yemen used Anthropic’s Claude to help develop guidance software for rockets and missiles, including a ballistic missile with a stated range of more than 2,000 kilometers and a system involving a hypersonic glide vehicle, according to the AI company.

Anthropic did not identify the group as the Houthis. The Financial Times said the militants were believed to be linked to the Iran-backed movement, which controls much of northern Yemen.

The case is one of six conventional-weapons operations Anthropic said it disrupted over the past year in China, Russia and Yemen. It offers a rare look at how an armed group used a commercial AI coding system for work that previously required specialist engineering teams.

The Yemen cell was working on three projects: a guided rocket using a phone-class flight computer and final-phase homing; a multi-stage ballistic missile with a range goal above 2,000 kilometers; and a family of missiles known internally as R2000 that included a hypersonic-glide variant.

The group used Claude Code for guidance, navigation and control software, which determines how a missile or rocket steers and remains stable in flight.

Anthropic said the users had Claude integrate an open-source autopilot with a low-cost flight computer, write control and position-estimation software, tune settings, build firmware and run flight simulations.

Several Claude instances were used at the same time. One wrote code, another conducted research and a third reviewed the work, effectively replacing parts of a small software-engineering team.

Anthropic’s safeguards stopped many of the requests, but not all.

The users concealed what they were building and divided work among separate sessions so that no single conversation disclosed the full purpose of the project, the company said.

The group went as far as conducting a live test of a guided rocket. Anthropic said the test appeared to fail. Within hours, the users returned to Claude to analyze what had gone wrong.

Anthropic said it had no evidence that the cell succeeded in fielding an operational weapon.

It banned accounts associated with the operation and shared information with public and private partners. But by then, the cell had created an offline simulation system that could operate without Claude or engineering software such as MATLAB.

The disclosure comes as the Houthis are again expanding their military position around the Red Sea. The group has seized Mocha and Mayun Island near the Bab el-Mandeb Strait, raising fresh concerns over shipping through one of the world’s busiest maritime chokepoints. Iran has long supplied the Houthis with weapons and technical support, although Tehran denies directing the group’s military operations.

Anthropic’s findings do not show that Claude designed an entire missile or gave the Yemen cell capabilities it previously lacked. The company said the actors already had access to weapons hardware and relevant technical expertise.

What the model supplied was software-engineering capacity.

That distinction matters because Anthropic’s separate tests found that newer AI systems can now perform some military and intelligence tasks that historically required scarce, highly trained specialists.

Its researchers tested models on activities including identifying targets from incomplete information and improving the performance of weapons systems. Anthropic said frontier models showed steady gains, while open-weight Chinese models, although less capable, also demonstrated potentially useful military skills.

The Yemen operation was not the only weapons case uncovered by the company.

Anthropic said Claude was also used by China-based actors to develop software for electronic warfare and suppression of enemy air defenses, work on an anti-torpedo system and collect intelligence on directed-energy weapons. Other actors used the model for Russian defense procurement and weapons-related intelligence gathering.

The company said it has introduced new classifiers intended to detect and block requests involving weapons development and high-yield explosives.

Its September threat report covers malicious activity detected between December 2025 and August 2026, ranging from weapons and cyber operations to surveillance, biological research, fraud and influence campaigns. The Yemen case stands out because the work moved beyond research and simulation to a physical weapons test.