China used US artificial intelligence to plan attacks on Taiwan

According to Anthropic, Chinese operatives have used its Anthropic Claude AI model for two military projects, two surveillance operations, and extensive knowledge acquisition. One military actor used Anthropic Claude to develop specifications for an anti-torpedo fire control system, test it against U.S. Navy systems, and produce a comprehensive technical proposal. Another researcher affiliated with the Chinese military created 16 electronic warfare and air defense suppression modules, with the given scenario involving 12 targets in Taiwan, such as Patriot and Tien Kung air defense systems and radar stations.

In addition to military applications, Anthropic has disrupted surveillance operations targeting Uyghurs outside China. An actor with ties to Chinese authorities posed as an Arabic-speaking expert. Anthropic’s Claude model helped analyze data from more than a hundred WhatsApp groups and Telegram channels, map social networks, and target exiled journalists and activists.

The biggest abuse is industrial data mining, or model distillation. Between May and July 2026, the Chinese company Alibaba conducted more than 151 million queries through thousands of fake accounts to use the acquired knowledge to train its own Qwen 3.x model in the areas of programming and logical reasoning. Similar methods are also used by DeepSeek, Xiaomi, and Zhipu/Z.ai, according to the manufacturer, Anthropic.

Such extensive use of the American Anthropic Claude model by Chinese actors has raised speculation about the true independent capabilities of the Chinese AI industry. For the most demanding tasks, they still rely on foreign solutions.