Summary
- China's Cyberspace Administration has summoned seven firms mentioned in Anthropic's report, focusing its investigation on DeepSeek and Moonshot.
- Claims suggest that Moonshot directed over 23 million exchanges to Claude via 5,380 fake accounts, while DeepSeek accounted for 12.1 million exchanges in just two weeks in July.
- The timing is critical for both firms: DeepSeek is briefing the UN Security Council on AI risks this week, while Moonshot is preparing for a $3 billion IPO in Hong Kong.
China is utilizing Anthropic's own leak detection methods against two of its leading AI companies, DeepSeek and Moonshot AI, amid allegations of routing sensitive user data to Anthropic's servers, as reported by The Information.
Officials from the Cyberspace Administration of China have reportedly visited the offices of both companies to conduct interviews with executives and staff. The focus of the investigation is whether sensitive information related to police, military, or state-affiliated corporations was transferred to servers in the U.S.
Throughout 2026, Anthropic has accused various Chinese labs of stealing outputs from Claude to create cheaper alternatives. They alleged that these startups were redirecting their queries to Anthropic, effectively providing users with Claude's responses at a lower cost.
Now, Beijing is examining the validity of those allegations.
The investigation was prompted by Anthropic's fourth threat intelligence report, a comprehensive 154-page document released on September 10, which accused seven labs in China—Alibaba, Moonshot, DeepSeek, Zhipu (Z.ai), MiniMax, SenseTime, and Xiaomi—of "illicit distillation."
In AI terminology, distillation refers to the practice of training a smaller model using the outputs of a larger one. Anthropic acknowledges that this practice can be legitimate but objects to it being done through fraudulent means.
The CAC initially summoned all seven companies mentioned in the report but has since focused its investigation on DeepSeek and Moonshot, with no penalties announced yet.
Both companies face challenging circumstances. DeepSeek is set to present to the United Nations Security Council about AI risks this week, sharing the platform with Anthropic CEO Dario Amodei. This event coincides with President Donald Trump's summit with Xi Jinping on September 24, where AI is expected to be discussed.
Moonshot is also in a precarious position, having confidentially filed for a $3 billion IPO in Hong Kong on September 3, aiming for a valuation of $50 billion. Any regulatory investigation findings would need to be disclosed in its prospectus prior to the IPO proceeding.