Chinese AI labs used over 151 million Claude exchanges for unauthorized training, Anthropic says
Anthropic reports that Chinese companies like Alibaba, Moonshot, and DeepSeek conducted illicit data extraction campaigns using its AI models. The largest campaign involved more than 151 million user exchanges between May and July.
Anthropic announced Thursday that it disrupted unauthorized attempts by several Chinese artificial intelligence laboratories to train their models without permission. The U.S. company stated these groups engaged in what it termed illicit distillation, utilizing data outputs from Claude to replicate capabilities of its system. A report released on Thursday details the specific actions taken against China-based firms including Alibaba, Moonshot, and DeepSeek. Anthropic confirmed that some of the intercepted exchanges contained sensitive information from individual users, major multinational corporations, and state-affiliated entities. The investigation noted activities originated between January and July, spanning locations in China, Iran, and West Africa. Operators affiliated with Alibaba were accused of using Claude outputs to assist in training its Qwen models. The scale of the operation involving Alibaba exceeded all previously measured campaigns at the company, totaling more than 151 million exchanges during the period from May through July. Daily activity reached nearly three million transactions originating from over 3,500 fraudulent accounts. Moonshot routed specific user requests to Claude and utilized the resulting data for model training, while DeepSeek was also implicated in similar unauthorized efforts.
