China Investigates DeepSeek Over Alleged Secret Routing of User Data

China’s internet regulator is reportedly investigating DeepSeek over allegations concerning how the Chinese AI company handles user data and interacts with foreign artificial intelligence models. The investigation follows claims by Anthropic, the company behind Claude, that DeepSeek had been routing some requests from its users to Claude without informing them. The allegations have raised questions about data protection, privacy and the methods used by AI companies to improve their models.

According to Anthropic, DeepSeek allegedly used a system that redirected selected user requests to Claude models and collected the resulting exchanges. Anthropic said the activity was connected to efforts to extract capabilities from its models through a technique known as AI distillation, in which outputs from a more capable model can be used to help train another system. Anthropic said it had identified and disrupted several such campaigns involving Chinese AI companies.

Anthropic further alleged that some of the information passing through DeepSeek’s systems could have included sensitive material supplied by users. The company said examples included internal corporate documents, software credentials and information connected to government-related activities. Anthropic’s report said DeepSeek had conducted more than 12 million exchanges associated with the alleged distillation activity over a 14-day period in July. These figures and allegations come from Anthropic’s investigation and have not been independently established by Chinese authorities.

The reported Chinese investigation also involves Moonshot AI, the company behind the Kimi AI model. Anthropic has made similar allegations against Moonshot, saying it observed millions of exchanges that it believes were connected to efforts to obtain capabilities from Claude. The reports have brought increased attention to the growing competition between Chinese and US AI companies and the methods being used to develop increasingly capable models.

The investigation comes shortly after US security agencies also accused several Chinese AI companies, including DeepSeek, Moonshot and Alibaba, of systematically extracting capabilities from American AI models. The US agencies said such activity had been occurring at significant scale, while Chinese companies have faced increasing scrutiny over their approaches to AI development. DeepSeek is meanwhile expected to participate in a United Nations Security Council discussion on AI and international security this week, placing the company at the centre of a wider global debate over AI development, security and governance.