Chinese AI companies allegedly copied US models by quietly sending their own users' questions to Anthropic — and ended up leaking sensitive Chinese military and corporate data to an American firm in the process.

• Anthropic says Moonshot AI secretly forwarded customer requests to Claude, then showed Claude's answers back to users as its own.
• The scale was large: close to 300,000 forwarded requests in a ten-day stretch, and more than 23 million interactions tied to the practice between May and July.
• One user believed to be linked to China's military handed over camera footage from hundreds of surveillance cameras in Chengdu — including cameras near military sites and a state defense research institute — which then landed in Anthropic's systems.
• An engineer at a big Chinese state-owned company also exposed internal code and working access keys for several major Chinese firms the same way.
• DeepSeek is accused of doing something similar, with leaked material including Chinese police surveillance systems and login credentials for a Russian defense ministry database.

Outlook: Expect tighter scrutiny of Chinese AI providers over both data handling and unauthorized use of US models, and more pressure on Beijing to police what its own firms send abroad.