Anthropic report alleges Chinese-linked actors used Claude for military and surveillance projects
Anthropic's September 2026 threat report claims China-linked agents employed its Claude model for at least five programs, including two military efforts. One actor drafted specifications for an anti-torpedo fire-control system and tested it against U.S. Navy systems, while another developed 16 electronic-warfare modules with scenarios involving Taiwan. The report also notes a program to distill Claude's capabilities, with 151 million training queries routed to Alibaba.
The report identifies five distinct programs linked to China-based actors using Claude. Two focused on military applications: one drafted specifications for an anti-torpedo fire-control system and tested it against U.S. Navy systems, while another produced 16 electronic-warfare modules with scenarios targeting Taiwan. Two additional programs involved surveillance work, and a fifth sought to distill Claude's capabilities, routing 151 million training queries to Alibaba.
Anthropic noted that account metadata and safeguards-flagged content pointed to PRC research institutions, including the PLA Academy of Military Sciences. The report suggests actors may have chosen Claude because U.S. frontier models trained on English-language material offer extensive knowledge of American military systems, useful for coding and reasoning workflows.
This report could heighten concerns about U.S. frontier AI models being exploited by foreign military actors, potentially accelerating debates over export controls and model access restrictions. Defense contractors and policymakers may face pressure to reassess AI supply chains, while the broader public could see increased scrutiny of AI safety measures. The alleged use of Claude for Taiwan-related scenarios may also amplify geopolitical tensions, though the report's findings remain allegations at this stage.