Houthis used Anthropic AI to try to build ballistic missiles

Anthropic has disclosed that a cell of militants in northern Yemen, a region controlled by the Houthi rebels, used its Claude AI model to develop guidance, navigation, and control software for ballistic missiles and guided rockets. The disclosure, contained in the company’s intelligence threat report released Thursday, highlights the growing risks of advanced AI falling into the wrong hands. While Anthropic said its safeguards blocked many of the requests, it acknowledged that not all were stopped, and the militants employed evasion tactics that exposed the limits of current AI safety measures.

The cell was running three weapons development programs, according to the report. The targets included a multi-stage ballistic missile, a multi-variant missile, and a guided rocket built around a “commodity phone-class flight computer.” In each case, the militants used Claude Code, Anthropic’s agentic coding tool, in place of human software engineers. The AI was used to integrate an open-source autopilot onto the phone-class computer, write control and position estimation software, tune control settings, run a firmware build pipeline, and perform flight simulations. A missile test by the group apparently failed, but the episode demonstrates that non-state actors with limited engineering capacity can now access sophisticated software development assistance.

Anthropic said the actors hid their goals and the intended use of the software, and split their work across multiple sessions so that no single session revealed their full intent. This piecemeal approach allowed them to circumvent safeguards that operate on a request-by-request basis. The report also noted that scientists made multiple attempts to use Claude to research biological weapons, underscoring a broader pattern of attempted misuse. The company did not name the Houthis, but the geographic reference to northern Yemen, combined with the group’s known missile programs and Iranian support, leaves little ambiguity about the actors involved.

The disclosure arrives at a moment of heightened regional tension. The Houthis, designated a terrorist group by the United States and a key member of Iran’s axis of resistance, this week launched an offensive along Yemen’s Red Sea coast, capturing a strategic port and advancing toward the Bab al-Mandeb Strait, a vital chokepoint for international shipping. The advance helped push Brent crude above $105 a barrel. The group, which has controlled Sana’a and much of Yemen’s populous north since 2014, has been steadily developing its missile capability with weapons and technology supplied by Iran. The renewed offensive deals a significant blow to Saudi Arabia, which has led an Arab coalition against the Houthis since 2015, and strengthens Iran’s hand as it seeks to pressure energy markets.

The implications extend beyond the immediate conflict. Frontier AI labs face a constant struggle to prevent unauthorized actors from using their models to attack the United States and its allies. The Yemeni case shows that safety measures, while partially effective, can be defeated by determined adversaries who understand how the systems work. The use of a commodity phone-class flight computer