Yemeni Cell Used Anthropic's Claude 'In Place of Human Software Engineers' To Develop Missile Guidance Systems

A weapons-development cell operating from Houthi-controlled northern Yemen used Anthropic's Claude AI 'in place of human software engineers' while developing guidance systems for rockets and missiles, according to the company's latest threat intelligence report.

Anthropic said the group was working on three weapons programmes and used Claude Code to help produce guidance, navigation and control software, the systems responsible for steering and stabilising a weapon in flight.

The company did not explicitly identify the actors as Yemen's Iran-backed Houthi movement. However, the cell operated in northern Yemen, territory largely controlled by the Houthis, and outside reporting has described the actors as likely linked to the group.

Claude Used Like a Team of Engineers

Anthropic said the actors ran several Claude instances simultaneously and assigned each one a different job. 'The actors used Claude Code in place of human software engineers,' the company said.

One instance would produce software, another would conduct research, and a third could review the work, resembling how a lead engineer might divide tasks among a small development team.

The projects included a guided rocket, a multi-stage ballistic missile with a stated range goal exceeding 2,000 kilometres and a separate missile family that included a proposed hypersonic glide vehicle.

Anthropic said Claude was used for tasks connected to flight-control software, simulations and troubleshooting. The incident highlights a growing concern for AI companies: increasingly capable coding models can perform technical work that once required teams of specialised engineers.

Safeguards Failed To Stop Every Request

Anthropic acknowledged that its safety systems blocked many of the cell's requests, but not all. According to the company, the users concealed their ultimate objectives and divided the project across multiple conversations, making it harder for any single session to reveal the full weapons-development effort.

Anthropic eventually detected the activity through an internal investigation into suspected weapons development and banned the accounts involved. It also shared intelligence about the operation with unnamed public and private-sector partners.

Guided Rocket Test Apparently Failed

The company said it found no evidence that the cell successfully fielded an operational weapon developed with Claude. However, there was a real-world test.

The actors test-fired a guided rocket in Yemen, according to Anthropic, but the launch apparently failed. Within hours, they returned to Claude and attempted to determine what had gone wrong.

Anthropic also found evidence that the group had developed an offline simulation toolkit, meaning parts of its engineering work no longer depended on Claude or specialist software such as MATLAB.

Anthropic Warns of Growing AI Weapons Risk

The Yemen case formed part of a wider September report documenting six conventional-weapons investigations involving actors in Yemen, China and Russia.

Other cases involved software for drone swarms, electronic warfare, military procurement and weapons research. Anthropic said it had disrupted the operations and used what it learned to strengthen its safeguards.

The company has also developed new evaluations to measure whether advanced AI models can perform military and intelligence tasks that traditionally require scarce human expertise.

Anthropic said those tests show frontier models are making 'consistent progress' in such areas. For the Yemen cell, that risk was no longer purely theoretical: according to Anthropic, Claude had already become part of an active effort to develop and test guided weapons.