The American artificial intelligence company Anthropic has disclosed that a cell of threat actors based in northern Yemen used its Claude AI model in an attempt to develop guided rockets and missiles, raising fresh concerns about the military misuse of large language models. The company published its findings on Thursday in a report documenting how it had detected and disrupted illegal uses of Claude across several countries.

According to Anthropic's report, the cell was running three distinct weapons development programmes: a guided rocket using a smartphone-class flight computer with final-phase homing guidance; a multi-stage ballistic missile with a stated range goal exceeding 2,000 kilometres; and a missile incorporating a hypersonic glide vehicle variant. The group used Claude as a substitute for a team of specialist engineers, leveraging it to develop guidance, navigation and control software for airborne weapons. Anthropic said it found no evidence that the actors succeeded in fielding an operational device, though they did conduct a live test-fire of a guided rocket that appears to have failed — and returned to Claude within hours to diagnose why. The company said it banned the associated accounts as soon as the illegal activity was detected and shared intelligence with public and private partners to reduce further risk.

Anthropics did not formally identify the cell, but the group's location in northern Yemen — a region largely controlled by the Houthi movement, an Iran-backed rebel force that holds much of the country including the capital, Sanaa — makes Houthi involvement the most likely explanation. The disclosures come as Houthi forces have launched a broad offensive, advancing toward the Bab al-Mandab strait, the narrow waterway connecting the Red Sea to the Gulf of Aden that serves as a critical shipping route between Europe and Asia. On Friday, the Houthis claimed control of an island in the middle of that strait. Yemen, a country on the southwestern tip of the Arabian Peninsula, has been embroiled in conflict for years; after a period of relative calm, fighting resumed in July amid a broader regional escalation.

The weapons-related case was not the only abuse flagged in Anthropic's report. The company also said it had disrupted attempts to use Claude models for biological weapons development and uncovered a suspected Russia-linked cyber espionage campaign targeting Ukraine. It additionally accused Chinese competitors of mounting hacking attempts aimed at extracting Claude's underlying capabilities.

The revelations underscore growing pressure on AI developers to prevent their tools from being weaponised. The dual-use risk of powerful language models — capable of synthesising engineering expertise that would otherwise require large, specialised teams — has become a central concern for governments and researchers alike. Separately, two Anthropic researchers warned this week that rapidly advancing AI poses an existential risk to humanity, adding to the urgency of the debate around safety and oversight in the industry.