Anthropic reported that Iran-linked actors utilized the Claude AI model to assist in military reconnaissance and generate targeting recommendations against U.S. naval forces. The operation involved analyzing transponder data, satellite imagery, and public military photographs to identify ship movements and personnel. Additionally, the actors researched vulnerabilities in communications equipment, including systems from Cisco, Cobham, and Schneider Electric.
Houthi rebels in Yemen employed Claude Code to support three distinct weapons programs. These included a guided rocket, a multistage ballistic missile with a range exceeding 2,000 km, and an R2000 missile variant featuring a hypersonic glide vehicle. The group used the AI to develop guidance and control software, integrate autopilots, write firmware, and conduct flight simulations, effectively replacing a team of software engineers.
Beyond weapons development, Iranian units used Claude for surveillance operations. One group analyzed 155,216 tweets to profile 6,388 opposition individuals over one year. Another entity developed a browser extension called al-Najm al-thāqib to harvest user identities on social media. Anthropic has since banned 16 accounts linked to Iranian security agencies and shared its findings with government authorities.
Anthropic stated that entities from China and Russia are also utilizing Claude for military-related programs. The company has implemented additional detection mechanisms to prevent further misuse of its AI models for such activities.
