China's Cyberspace Administration Investigates DeepSeek and Moonshot AI Over Alleged Data Leaks to Anthropic
nashnova research
China's internet regulator is investigating DeepSeek and Moonshot AI after Anthropic accused both of secretly routing tens of millions of user conversations to its Claude model — raising the prospect that sensitive Chinese military and police data reached American servers.
What actually happened?
Anthropic published a 154-page report on September 10, accusing seven Chinese companies — including DeepSeek and Moonshot AI — of large-scale unauthorized use of its Claude model.
The mechanism: when users queried DeepSeek or Moonshot's models, the companies secretly forwarded those requests to Claude behind the scenes, then returned Claude's answers as if they were their own models' output.
This means → users believed they were talking to a domestic Chinese AI, while their data was crossing the border to Anthropic's servers in the United States.
How much data was involved?
Moonshot AI forwarded more than 23 million conversations to Claude between May and July this year.
DeepSeek forwarded over 12 million in just 14 days in July.
In plain terms = the two companies quietly sent more than 35 million user conversations to an American AI system within a few months.
Why does this involve military and police data?
The report revealed that DeepSeek forwarded an engineer's request to build a police surveillance system to Claude.
A more serious case: a user assessed as likely affiliated with the People's Liberation Army used Moonshot's Kimi model to analyze surveillance footage from hundreds of police cameras in Chengdu — including cameras outside PLA facilities and defense-related agencies. Moonshot forwarded the request and the footage to Claude without the user's knowledge.
This reflects something beyond commercial misconduct — defense-related data may have reached an American company's servers.
What is each side worried about?
The U.S. side is focused on Chinese developers using "distillation" — training their own models on another model's outputs — to exploit American AI systems without authorization.
The Chinese side faces a thornier problem: the forwarding may have violated China's data-security laws. China tightly controls cross-border data transfers; companies sending large volumes of data involving Chinese citizens and businesses abroad must obtain CAC approval.
In plain terms = the U.S. worries about stolen technology; China worries about exported data. The same act creates problems on both ends.
Where does the investigation stand?
After the Anthropic report dropped, the CAC summoned representatives of all seven named companies, then narrowed its investigation to DeepSeek and Moonshot AI, sending officials to both offices to interview executives and staff.
The other five companies named are Alibaba, Zhipu, SenseTime, MiniMax, and Xiaomi.
The investigation is ongoing. The CAC has not yet decided whether to impose penalties and is assessing the severity of the violations and the companies' intent.
Why is the timing so sensitive?
The probe comes as President Xi Jinping visits Washington this week, expected to discuss establishing an "AI dialogue" mechanism with President Trump.
The mechanism would let both countries notify each other of AI incidents with national-security implications.
This means → the investigation's outcome will test the real boundaries of data compliance for China's AI industry. Too light a penalty undermines Beijing's credibility with Washington; too harsh a penalty risks chilling its own AI sector.
市场有风险，内容仅供研究参考，不构成投资建议。