Did a failed missile test alert Anthropic to Iran-backed Houthis' use of Claude?

Iran-backed Houthi militants in Yemen may have used Claude, Anthropic’s AI chatbot, to develop software for guided missiles, a threat intelligence report published by the tech company indicates.

The 154-page report, published on Thursday, details how governments, criminal groups and others misused Anthropic's powerful AI models in an attempt to develop biological weapons, conduct espionage, disseminate propaganda, and other malicious activities.

One startling revelation was how a Yemen-based group used Claude Code to work on guidance, navigation and control software for a guided rocket and other missile programmes. While Anthropic did not name the group involved, details in the report point to the Houthis, who have built a sizeable arsenal of missiles and drones.

Anthropic said these "threat actors" were also developing a ballistic missile with a stated range of more than 2,000 km and a missile variant featuring a hypersonic glide vehicle. The company said it did not have evidence that the actors successfully deployed an operational guided weapon, though they did test-fire a guided rocket that apparently failed.

AI USED AS ENGINEERING TEAM

According to Anthropic, a weapons engineering cell was based in northern Yemen and working on three missile programmes. One involved a guided rocket using a commercially available, phone-class flight computer and final-phase homing guidance.

Another was a multistage ballistic missile with a stated range of more than 2,000 kilometres. The third, described as the “R2000” set, included several missile variants, including a hypersonic glide vehicle.

The militants used Claude Code to develop guidance, navigation and control software, or GNC, which helps steer and stabilise a flying vehicle.

Anthropic said Claude was used to integrate an open-source autopilot with a phone-class flight computer. The AI helped write control and position-estimation software, tune control settings, run firmware builds and conduct flight simulations.

The operators also ran several Claude instances simultaneously, assigning different tasks to each. One instance was used to write code, another to conduct research and a third to review the first one's work.

Anthropic said its safeguards blocked many requests but couldn't stop all of them. "The actors used a variety of tactics to evade our safeguards, including hiding their goals and the products the software was meant for, and they split their work across multiple sessions, so no single session revealed their full intent," it said.

Within hours of the group's guided missile test-fire ending in failure, they returned to Claude to troubleshoot.

Anthropic said it identified the activity during an internal investigation, banned accounts linked to the actors and shared information with public- and private-sector partners.

WIDER MISUSE

The Yemeni case was one of six examples in which Anthropic identified attempts to use Claude for conventional weapons development. The other cases involved actors in China and Russia and included work on firearms, missiles, armed drones, bombs and other munitions.

The report also highlighted the use of Claude in cyber operations, surveillance and propaganda. Anthropic cited Russian espionage and "smash-and-grab" cyberattacks, as well as China-linked surveillance programmes targeting Uyghurs in Syria and domestic dissidents.

It also documented propaganda-related activity involving Russia, Malaysia, Iran and Bangladesh.

Anthropic said some of its most concerning cases involved biological research. It identified five instances in which scientists used its AI models for research that could potentially assist in developing biological weapons.

The company said it did not disclose the researchers, institutions, countries or specific biological agents involved because it remained uncertain about their intentions. However, the findings underline the growing concerns among global security experts about terrorist groups using increasingly capable AI systems for weapons development.

"These cases are not typical: we’re highlighting some of the most sophisticated misuse we’ve seen," Anthropic said. "But they’re especially important to discuss, because they show us where AI misuse is headed, where our safeguards work, and where they need to improve."

The report comes days after Anthropic researcher Jacob Coxon quit and cracked open a veritable can of worms. He stated that AI is moving too quickly and will one day outpace humans’ ability to control it. It could build "extinction-level bioweapons" or "kill us all by the end of the decade", he claimed.