Between December 2025 and August 2026, Anthropic disrupted what it calls the largest AI model theft it has ever measured. Seven Chinese AI companies used thousands of fake accounts to pump user questions into Claude, harvest the responses, and feed them into their own training pipelines.
The scale is staggering
Alibaba alone generated 151 million exchanges through 3,500 fraudulent accounts between May and July, peaking at nearly 3 million prompts per day. Moonshot and DeepSeek went further: they routed their own customers' requests to Claude without telling anyone, displayed Claude's answers as if they came from Kimi or DeepSeek, then kept the transcripts for training.
That means if you asked Kimi a question in July, Claude may have answered it—and both companies now own that exchange. Anthropic says the forwarded prompts included names, emails, and corporate data from hundreds of users, likely violating privacy law in a dozen jurisdictions.
Why it matters
This is not espionage. It is wage theft at the infrastructure layer. You pay DeepSeek. DeepSeek pays nothing. Claude does the work. DeepSeek keeps the margin and the training data. The fact that it is widespread, industrial-scale, and involves some of the biggest names in Chinese AI suggests this is strategic, not rogue.