China Probes DeepSeek and Moonshot Over 23 Million User Data Exchanges Sent to Anthropic Claude

China Turns Investigation Inward After Anthropic's Distillation Claims

China's Cyberspace Administration (CAC) has launched an investigation into DeepSeek and Moonshot AI over user data that may have reached Anthropic's servers without customer knowledge1

1

2

2

. The China probe was triggered by Anthropic's 154-page threat intelligence report published on September 10, which accused seven China-based AI labs of routing customer requests through Anthropic Claude to harvest outputs for model training—a practice the company calls illicit distillation3

3

. Officials from China's Cyberspace Administration have visited both firms to question executives and staff, with the investigation ongoing and no penalties yet determined4

4

.

The named companies initially included Alibaba, Moonshot AI, DeepSeek, Zhipu, MiniMax, Xiaomi, and SenseTime. While the CAC summoned representatives from all seven firms, the investigation into AI companies quickly narrowed its focus to DeepSeek and Moonshot due to specific examples detailed in Anthropic's report1

1

. The regulatory scrutiny centers on whether sensitive information—including police surveillance footage, military-related data, and state-linked corporate information—ended up on U.S. servers through data routing to Claude2

2

.

Massive Scale of Alleged Data Routing to Claude

The figures in Anthropic's report reveal the substantial scale of alleged data transfers. Anthropic counted more than 23 million exchanges from Moonshot between May and July, and almost 300,000 requests in a single ten-day window, run through 5,380 fraudulent accounts1

1

. DeepSeek allegedly routed more than 12.1 million exchanges to Claude during a 14-day period in July5

5

. According to Anthropic, Moonshot forwarded requests silently and showed the answers as though its own Kimi model had produced them, effectively passing off Claude's outputs as original work1

1

.

The user data sent to Anthropic reportedly included names, email addresses, company data, and other sensitive data of hundreds of end users in at least a dozen languages2

2

. One particularly concerning example involved a user assessed as likely affiliated with the People's Liberation Army who asked Moonshot's Kimi to analyze surveillance footage following a person across hundreds of police cameras in Chengdu, including cameras outside PLA facilities and defense-linked institutes1

1

. This type of sensitive data falls under categories that China's cross-border data rules treat most carefully, explaining why the probe narrowed to these two companies.

Two Different Complaints From the Same Evidence

Anthropicโ€™s original grievance focused on what left Claude—Chinese labs taking its model's output and passing it off as their own through model distillation. Beijing's concern represents the opposite perspective: Chinese users' data landing on an American company's servers without those users knowing1

1

. This distinction matters because it means China is not endorsing Anthropic's complaint but has found its own violation within the same evidence. The practice Anthropic describes as illicit distillation involves training a smaller model on a bigger one's outputs—legitimate on its own, but objectionable when done through fraudulent accounts4

4

.

The data leaks to Anthropic's Claude raise questions about data privacy and whether the practices violated privacy laws and the labs' own terms of service2

2

. For Chinese regulators, the issue centers on unauthorized data transfers abroad, particularly when involving police, military, and state-affiliated information. This creates a domestic regulatory problem before it becomes an international intellectual property dispute.

Timing Creates Complications for Both Companies

The timing of the investigation creates significant complications for DeepSeek and Moonshot. DeepSeek is scheduled to brief the UN Security Council on AI risks this week, presenting itself as a responsible voice on international security while its own regulator examines its data handling practices in Beijing1

1

4

4

. The session lands just a day before President Trump's summit with Xi Jinping on September 24, where AI is reportedly on the agenda and US-China tensions over artificial intelligence are expected to feature prominently4

4

.

Moonshot faces a different challenge with its Hong Kong IPO plans. The company confidentially filed for a Hong Kong IPO on September 3, targeting $3 billion at a $50 billion valuation4

4

. An open regulatory investigation would need to appear in the prospectus before the listing can proceed, potentially complicating the timeline and investor confidence. Neither DeepSeek nor Moonshot has commented publicly on the probe1

1

.

Broader Context of AI Competition and Regulation

The dispute adds another dimension to growing US-China tensions over artificial intelligence and intellectual property5

5

. U.S. agencies recently warned American AI companies about alleged large-scale efforts targeting proprietary model capabilities. Treasury Secretary Scott Bessent previously indicated that sanctions could follow if evidence verified Chinese model distillation, stating the administration would be "looking at that in the coming days or weeks"5

5

.

Interestingly, on September 10—the same day Anthropic published its report—China rejected a U.S. intelligence advisory on Chinese AI distillation as unfounded, with the foreign ministry dismissing the accusations publicly1

1

. Twelve days later, its internet regulator had summoned every company Anthropic named. While these are separate documents, the timing highlights how quickly Beijing moved from public dismissal to internal investigation.

Zhipu, another company named in Anthropic's report, is simultaneously dealing with a separate data privacy issue after a developer found its ZCode tool uploading encrypted snapshots of local code repositories without consent1

1

. The company apologized, open-sourced the tool, and commissioned outside assessments. While unconnected to the Claude routing allegations, both incidents raise the same fundamental question: when Chinese AI products move user data somewhere unexpected, who can verify what happened to it and whether users' rights were protected? Any confirmed violations could influence privacy controls, model access rules, and safeguards around sensitive prompts across the industry5

5

.