Anthropic reveals Iranian actors used Claude to target US Navy warships

Anthropic's September 2026 threat report reveals that state-backed and freelance actors from China, Russia, Iran, and Yemen used its Claude AI models to advance military and cyber capabilities. Notable misuses include a Yemeni cell designing guided missiles, Russian freelancers building autonomous lethal drone swarms trained on Ukrainian combat footage, and an Iranian-linked operator compiling targeting data on U.S. Navy warships. Anthropic has banned the associated accounts and shared findings with government authorities.

Claude misuse by state actors

• ▪Anthropic released a threat intelligence report on September 10, 2026, detailing how state-backed and freelance actors used its Claude AI models for weapons development, cyber operations, surveillance, and fraud
• ▪The Yemen-based actors test-fired a guided rocket developed with Claude's assistance, and within hours of the test failing, returned to the AI model to diagnose the failure
• ▪A cell of threat actors in northern Yemen used Claude Code to develop guidance, navigation, and control software for a guided rocket, a planned 2,000 km ballistic missile, and a hypersonic glide vehicle variant

Chinese military AI applications

• ▪A China-based actor linked to Chinese research institutions, including the PLA Academy of Military Sciences, used Claude to develop electronic-warfare and air-defense suppression software simulating 12 targets in Taiwan
• ▪A China-based defense-intelligence actor used Claude to research foreign high-power microwave weapons, identify component suppliers, trace supply chains, and draft restricted briefings for senior Chinese officials
• ▪A China-based actor associated with a defense manufacturer used Claude to write a 200-page technical proposal and fire-control software specifications for an anti-torpedo system intended for the Chinese navy

Iranian targeting of US Navy

• ▪The Iranian-linked operator directed Claude to compile software vulnerability research on shipboard systems, specifically targeting ship satellite, communication equipment, and industrial control products
• ▪An Iranian-linked operator used Claude to collect and analyze public data, including scraped military photo captions, transponder identifiers, and satellite-imagery scripts, to develop targeting recommendations against U.S. naval forces

Russian autonomous drone development

• ▪Freelance Russia-based actors used Claude to build core software for an autonomous swarm of first-person-view attack drones, including shared swarm memory, fault-tolerant coordination logic, and terminal guidance
• ▪The Russian actors trained their drone target-selection tool on scraped Ukrainian combat footage and used a fixed coordinate in the Donetsk area of Ukraine as a demonstration strike point
• ▪The Russian drone-swarm developers designed their platform for autonomous lethal engagement, enabling the onboard model to select targets and issue detonation commands without a human in the loop

AI-enabled cyber operations

• ▪A Russian-speaking espionage actor tracked as GTG-20006 used Claude-driven AI agents to continuously rewrite and recompile malware to evade detection by common antivirus security products
• ▪Chinese-speaking operators based in Changsha used Claude-driven workflows to search for software vulnerabilities, develop exploits, and conduct intrusions against approximately 50 organizations, including foreign government networks

Anthropic account bans

• ▪Anthropic identified and banned the accounts involved in the military and cyber operations, shared the information with government authorities, and developed new detections for future misuse
• ▪Anthropic banned 16 Claude accounts operated by two units associated with Iranian paramilitary and domestic security agencies, including one that maintained an identity-record database of Iranian nationals

Debatable claims

• ▪Militaries should ban the deployment of fully autonomous lethal weapons
• ▪AI companies should restrict access to frontier models to verified users
• ▪AI labs should temporarily pause the development of more powerful models