China's Regulator Probes DeepSeek and Moonshot for Secretly Routing User Prompts to Claude

Two of China's most prominent AI labs are now under state scrutiny at home, and the trigger came from an American rival. China's Cyberspace Administration has opened a probe into DeepSeek and Moonshot AI after Anthropic accused both of quietly funneling user prompts into Claude to train their own models, Gizmodo reported.

The accusation sits in Anthropic's September misuse report, "Detecting and countering misuse of AI." The company says it counted more than 23 million Moonshot conversations between May and July, and 12.1 million DeepSeek prompts across a 14-day stretch in July. Those are the numbers the wire headlines mostly skipped.

Anthropic goes further on what was inside those prompts. It says the traffic carried sensitive material, names, emails, and company data, drawn from individual users, large multinationals, and state-affiliated actors, spread across at least a dozen languages.

Regulators reportedly questioned several executives at Chinese labs before the focus tightened onto DeepSeek's open-source models and Moonshot's Kimi assistant. The scrutiny fits a pattern where distillation and data flows keep drawing regulatory heat, the same territory covered in CISA's warning on Chinese AI distillation.

It is worth being precise about what this is. The Cyberspace Administration has launched a regulatory probe. It has not published findings, announced a penalty, or confirmed Anthropic's tallies, and neither DeepSeek nor Moonshot has admitted routing prompts to Claude in any of the accounts so far. The volumes and the sensitivity claims remain Anthropic's allegations.

The timing carries its own weight. The probe surfaces ahead of expected Trump-Xi meetings where AI is likely on the agenda, and some coverage reads Beijing's move as a possible gesture on privacy enforcement rather than a purely domestic housekeeping exercise. China's willingness to put its own champions under the microscope, right as it presses its case abroad, echoes the posture it brought to DeepSeek's turn at a UN Security Council AI briefing.

For now the story is an investigation and a set of contested numbers, not a verdict.