Yemeni militants attempted to use Anthropic's AI model, Claude, to develop guidance, control, and navigation software for ballistic missiles, according to a report released by the US-based artificial intelligence company. The report also highlighted separate efforts by users to employ the AI model in researching biological weapons, raising concerns about the potential misuse of advanced AI technologies.

Anthropic, which did not explicitly name the Iranian-backed Houthi rebels, stated it had identified a cell of threat actors operating in northern Yemen—an area largely controlled by the Houthis—engaged in three weapons development programs. The group is believed to be one of the main factions in Yemen’s ongoing conflict and is designated a terrorist organization by the US.

The report detailed attempts by this Yemeni cell to develop multiple sophisticated weapons systems, including a multistage ballistic missile, a multi-variant missile, and a rocket using what Anthropic described as a “commodity phone-class flight computer.” The militants reportedly used Claude’s coding capabilities instead of human engineers to write software for the guidance, navigation, and control systems essential to missile operation.

Anthropic noted that while its safety safeguards blocked many of the requests, some managed to circumvent these protections by using various evasion tactics. The group said the militants conducted a test firing of a guided rocket, which appeared to fail. They then resumed consulting the AI to troubleshoot the failure.

Alongside missile development, the company revealed that various actors also tried to use Claude to research biological weapons. Five specific cases were cited where users circumvented controls and tried to conceal the true intent of their queries, some from countries restricted from accessing Anthropic’s models, including Russia, China, and Iran. One example involved a researcher from an unsupported region who spent weeks planning experiments related to avian influenza, although the AI’s safety filters limited this work to its weakest model.

Anthropic stressed that it could not definitively determine whether those involved in biological research intended malicious outcomes, noting that the same scientific information could be used for vaccine development.

The report underscored the increasing risks posed by the proliferation of advanced AI tools and the efforts by developers to prevent unauthorized use that could threaten global security, particularly with regard to nations or groups hostile to the US and its allies.

Meanwhile, the Houthis have recently intensified their military campaign along Yemen’s Red Sea coast, capturing a strategic port and making advances near the Bab al-Mandeb Strait, a critical chokepoint for international maritime trade.

While Anthropic declined to confirm the identity of the northern Yemen cell, it is widely believed to be linked to the Houthis, given their control over the region and ongoing military activities. The company called for a broader discussion within the AI industry and among governments to address the emerging risks of AI-assisted weaponization and biological threats.