China investigates DeepSeek and Moonshot AI for alleged transfer of data to the United States

He who thinks he takes is taken. For months, Anthropic has accused Chinese laboratories of training their models on Claude's responses, a practice that we detailed when the report was released. Beijing in turn opens an investigation at the end of September 2026.

There is nothing American about the target. The Chinese Internet regulator is interested in DeepSeek and Moonshot AI, two of the national flagships of artificial intelligence.

DeepSeek and Moonshot in the sights of the Chinese regulator

According to The Information, the Cyberspace Administration of China (CAC) first summoned the seven companies named by Anthropic. She then narrowed her investigations into two of them. Agents went to the offices of DeepSeek and Moonshot to question managers and employees. No sanctions have been announced at this stage.

The question asked by Beijing is surprising. Investigators want to know if sensitive data, from the police, the army or companies linked to the state, landed on American servers. The suspicion of technological theft to the detriment of Anthropic takes a back seat. For China, the real risk is the leak of data to the United States.

It all starts from the fourth threat intelligence report published by Anthropic on September 10, 2026. This 154-page document accuses Alibaba, Moonshot, DeepSeek, Zhipu, MiniMax, SenseTime and Xiaomi of illicit distillation. Distillation involves training a smaller model from the responses of a more powerful model. Anthropic does not dispute the technique itself. She denounces the use of false accounts to achieve this.

The volumes are dizzying. Moonshot would have sent more than 23 million exchanges to Claude between May and July, including around 300,000 requests sent in ten days by a network of 5,380 fraudulent accounts. DeepSeek reportedly generated more than 12.1 million exchanges in fourteen days in July. Anthropic goes further, since it affirms that certain laboratories relayed questions from their own users directly to Claude. Chinese Internet users received responses from an American AI without knowing it, and their data went with them.

DeepSeek and Moonshot caught in geopolitical calculations

The schedule complicates everything. Moonshot, the company behind the Kimi model, filed a confidential application for an IPO in Hong Kong on September 3. It is targeting $3 billion in fundraising for a valuation close to $50 billion. An ongoing regulatory investigation must be included in the prospectus before any listing.

DeepSeek was invited to speak before the UN Security Council on the risks of AI, on the same platform as Dario Amodei, the boss of Anthropic. Artificial intelligence was also on the agenda of the summit between Donald Trump and Xi Jinping in Washington. Beijing has every interest in showing that it controls its champions, especially when the accusation comes from an American company.

Anthropic also has its detractors. Several observers mocked his accusations, recalling that the great American models were themselves trained on texts collected without authorization.