Iran Leveraged AI for Ballistic Missiles and Assassination Lists, Report Finds

A new report from Anthropic, the developer of the AI model Claude, has revealed how hostile actors exploited the system for intelligence and terror purposes between late 2025 and August 2026. The findings indicate a significant shift in technological capabilities, enabling AI to orchestrate complex operations from data collection to execution.

Most concerning were the activities of a group of engineers in northern Yemen who used Claude as a complete substitute for human software engineers. This team attempted to develop a guided rocket, a ballistic missile capable of traveling over 2,000 kilometers, and a hypersonic glide vehicle. After a rocket test failed, the team quickly returned to the AI model for analysis, though Anthropic found no evidence these systems became operational.

Concurrently, an individual linked to Iran utilized the AI model to build an automated intelligence tool. This tool gathered and analyzed open-source information on hundreds of individuals in Israel and Jewish communities worldwide, aiming to enrich target lists and create profiles of government officials and NGO activists. Another Iranian operative used Claude to formulate operational recommendations for attacking U.S. Navy forces, tracking ships, identifying personnel, and seeking maritime vulnerabilities.

Within Iran, two security units operated 16 accounts to analyze tweets and profile thousands of citizens for repression and monitoring of regime opponents. One unit developed a malicious browser extension disguised as a prayer time application to secretly harvest social media user identities. The Iranian regime's propaganda mechanisms also employed Claude to create fake identities and distribution networks for "cognitive warfare," attributing false claims to Western research institutions.

The report also noted activities by the Iranian opposition group Mujahedin-e Khalq and UAE government actors who drafted false testimonies for UN institutions. Additionally, Anthropic identified an account from a company described as Israeli-Singaporean that attempted to build a platform for mapping users in Iran and the Gulf and generating fake identities. However, this project was detected in its experimental phase, and access was blocked before it could be activated.