Anthropic’s report says foreign adversaries used Claude to help build surveillance systems and military targeting tools

Anthropic’s September report describes AI supplying engineering and intelligence labor to hostile operations. These are Anthropic’s selected investigations; attribution varies, and prototypes do not establish battlefield use.

Inside the national-security cases:

• Russia — defense secrets: State-linked hackers stole drone-company emails and proprietary software. A separate, apparently freelance team developed autonomous attack-drone software, tested in simulation and on hardware; combat deployment was not established.
• China — targeting and repression: A defense researcher built software simulating attacks on Taiwan’s air defenses. Separately, a government-aligned operation profiled Uyghurs abroad, including family ties usable as leverage; recruitment outcomes remained unverified.
• Iran — ships and citizens: One actor compiled targeting recommendations against US naval forces. Separately, security-linked units shipped identity-harvesting software supporting domestic surveillance. Planning an attack is not evidence it happened. State-aligned propagandists also fabricated claims attributed to CSIS, Brookings and RAND.
• Yemen — beyond a chatbot: A weapons cell test-fired a guided rocket, apparently unsuccessfully, then sought Claude’s troubleshooting help.
• Mali — banning accounts cannot undo everything: A surveillance platform covering roughly 25 million SIMs ran locally; Anthropic’s account ban did not stop it.