Anthropic's Claude AI Weaponized by Iran, Houthis, and Other Adversaries, Report Shows

Anthropic's latest threat assessment, released in September 2026, documents extensive misuse of its Claude AI system by adversarial actors seeking to advance military capabilities and intelligence operations. The findings demonstrate that Iran's military apparatus, affiliated Houthi forces, and other geopolitical rivals have systematically deployed the American-built model to support weapons development, naval targeting, and domestic surveillance initiatives.

Military Targeting and Naval Reconnaissance

Among the most significant discoveries outlined in Anthropic's report is the deployment of Claude by an Iran-linked threat actor to conduct military reconnaissance and generate targeting data against U.S. Navy assets operating in the Middle East. The adversary leveraged Claude to synthesize publicly accessible transponder data from ships and aircraft, cross-reference commercial satellite imagery, track U.S. naval movement patterns, and extract identifying information about American military personnel from captions embedded in publicly shared military photographs. The threat actor also conducted research into potential security gaps affecting shipboard communications systems, specifically examining documented vulnerabilities in Cobham Sailor VSAT terminals, Cisco communications equipment, and Schneider Electric EcoStruxure systems. In response, Anthropic terminated the associated account, deployed enhanced detection protocols, and coordinated with government agencies to share intelligence about the incident.

Weapons Development Programs

A separate investigation identified a Houthi-controlled cell operating in northern Yemen that employed Claude Code to accelerate development across three distinct weapons initiatives. The programs encompassed a guided rocket system incorporating a phone-class flight computer for terminal guidance, a multistage ballistic missile designed to achieve ranges exceeding 2,000 kilometers, and variants of the R2000 missile family featuring a hypersonic glide vehicle component. The group utilized Claude to generate guidance, navigation, and control algorithms; integrate open-source autopilot software with phone-class flight computers; produce control and position-estimation routines; optimize system parameters; construct firmware; and execute flight simulation scenarios. Anthropic characterized the arrangement as a substitution of multiple Claude instances for a conventional software engineering team, encompassing coding, peer review, technical research, and simulation functions.

Although the Houthis themselves are not Iranian nationals, the organizational structure allows for knowledge transfer and collaborative development with allied partners who possess manufacturing capabilities to produce the resulting weapons systems.

Surveillance and Monitoring Operations

Beyond weapons development, Iranian entities have deployed Claude for intelligence gathering and population monitoring. One security-linked Iranian unit leveraged the model to process 155,216 tweets within a single year, using the analysis to identify, profile, and track 6,388 individuals identified as opposition figures. A separate Iranian group utilized Claude as part of an engineering workflow to construct domestic surveillance infrastructure, including the operationally deployed "al-Najm al-thāqib" Firefox extension, which performs large-scale harvesting of user identities from major social media platforms. Anthropic has terminated 16 Claude accounts connected to Iranian paramilitary and domestic security organizations, though the company acknowledges that additional compromised accounts may remain active.

Broader Adversarial Adoption

The misuse of Claude extends beyond Iranian and Houthi actors. Anthropic's assessment indicates that both China and Russia are actively incorporating Claude into their respective military development initiatives.