Anthropic Report Details Adversaries’ Military Misuse of Claude

Summary

Anthropic’s latest 154-page misuse report describes how commercially available Claude tools were used by malign actors across several countries for military and other harmful activities over an eight-month period. An Iran-linked actor used Claude to collect public data, track U.S. naval positions, compile targeting material and research vulnerabilities in maritime communications and industrial-control equipment. In Yemen, a weapons-development cell used Claude Code to write and test guidance, navigation and control software for guided rockets and missile programs, including a project with a stated range goal above 2,000 kilometers and a hypersonic-glide-vehicle variant. The actors divided tasks among multiple Claude instances and evaded safeguards by concealing their goals and spreading work across sessions; Anthropic said a guided-rocket test appeared to fail, and found no evidence that an operational device had been fielded. A China-based actor used Claude to draft an anti-torpedo fire-control specification, a proposal exceeding 200 pages, acquisition documents and software components, while another China-based operation built electronic-warfare software to analyze radars, air-defense sites and communications nodes, including scenarios involving targets in Taiwan. In Russia, a freelance group used Claude Code to develop software for an autonomous FPV drone swarm, including target recognition, coordination and detonation logic without a human in the loop; Anthropic could not confirm a direct government link. The company says it banned associated accounts, shared threat intelligence and added findings to its safeguards, but the cases show that geographic controls and weapons-use restrictions can still be bypassed. The report also places these incidents within wider debates about AI-enabled military operations, public-data exposure, autonomous weapons and the limits of commercial model guardrails.