16:40Machine-translated from Azerbaijani by AIRead in Azerbaijani
Claude Used for Military Projects in China
In a threat report released in September 2026, Anthropic disclosed that hundreds of people linked to China had used the Claude model in at least 5 projects. Of these projects, 2 were military, 2 involved surveillance, and 1 aimed to transfer the model’s capabilities. According to the report, Claude was used to develop weapons systems, analyze targets in Taiwan, and track Uyghurs living abroad. The company supports these claims with account information, the content of queries, and security system logs.
In one of the military projects, Claude was used to develop fire-control rules for an anti-torpedo defense weapon. Using information from open sources, the model tested the system by comparing it with the US Navy’s countermeasures against torpedoes and submarines. The person involved in the project presented himself as a manufacturer in the US defense sector and prepared a technical proposal of more than 200 pages. Anthropic believes the person is linked to a Chinese defense manufacturer.
Another military researcher used Claude to develop approximately 16 software modules intended to disable electronic warfare and air defense systems. The software analyzed radars, surface-to-air missile systems, command posts, and communications hubs, then ranked the targets by importance. One scenario identified 12 targets in Taiwan. The list included Patriot and Tien Kung batteries, air bases, an early-warning radar, and a command bunker. Account information showed that the person had ties to Chinese research institutions, including the People’s Liberation Army Academy of Military Sciences.
One of the surveillance operations aimed to infiltrate armed Uyghur groups in Syria and track Uyghur activists living abroad. The person involved in the operation presented himself as an Arabic-speaking specialist. Claude helped process information collected from more than 100 WhatsApp groups and dozens of Telegram channels. The model established links between people on different platforms, identified their social circles, and selected suitable individuals to recruit for cooperation.
People facing financial hardship, separated from their families, or politically dissatisfied were considered more vulnerable targets. People with relatives living in Xinjiang were also singled out. Claude drafted deceptive messages in local dialects, translated conversations in real time, and evaluated how convincing the messages to be sent would be. As part of the same operation, mass complaints were filed against Uyghur Post journalists, and bots were used to conduct information dissemination campaigns.
Anthropic also accused Chinese companies of transferring Claude’s capabilities to their own artificial intelligence models. In the largest campaign, Alibaba sent more than 151 million queries from May through July 2026. On some days, the number of queries sent from thousands of fake accounts approached 3 million. According to the company, the collected responses were used to teach the Qwen 3.x model logical analysis, programming, and the completion of long-running tasks.
The report also named DeepSeek, Xiaomi, and Zhipu in connection with similar activities. DeepSeek generated more than 12.1 million queries over 14 days. For Xiaomi, the figure was more than 400,000. Fake accounts, intermediary networks, and conversations collected from other users were used for this purpose. This method allows responses from a powerful model to be collected at scale and then transfers its knowledge and capabilities to another model using less time and computing power.