China probes DeepSeek and Moonshot over data

What China is investigating

I would not reduce this story to another dispute over model copying. According to The Information, China’s internet regulator is examining whether user data may have been sent abroad through Claude. As of September 24, 2026, this is an investigation, not a confirmed violation.

In its report on alleged unlawful distillation, Anthropic said Moonshot AI sent more than 23 million exchanges to Claude between May and July. For DeepSeek, the company cited more than 12.1 million exchanges over 14 days in July. Anthropic alleges the activity involved proxy accounts, intermediary networks, and stolen or forged credentials.

The mechanism is straightforward: a user believes they are talking to a Chinese model, but their prompt is silently forwarded to an external service. Claude’s response is then returned and, Anthropic alleges, enters the training pipeline of a competing model. Names, email addresses and corporate documents could have traveled with the prompt.

That is where this becomes more serious than ordinary distillation. The issue is no longer only where training data came from, but where a request was actually processed and whether the user knew about it.

• Anthropic is the primary source for the reported traffic volumes.
• The Information reports that a Chinese regulatory inquiry has begun.
• The available source material contains no final regulatory findings yet.

Why hidden routing changes the equation

The main risk is architectural: a promise to keep data within a country means little if an intermediary layer can send prompts to an external API. I would first examine routing logs, account provenance, the chain of intermediaries and the rules governing attachments.

The problem is not always obvious to users. The interface may belong to one provider, the model displayed on screen may carry its name, while actual inference is partly handled by a completely different system. This architecture undermines consent for data processing and control over corporate secrets.

• Providers may have to disclose where requests are processed more precisely.
• Regulators have reason to inspect not only storage, but also data transit.
• Model developers will find it harder to conceal external dependencies behind their own interface.

Millions of exchanges make the allegation substantial, but they do not prove it automatically. The most uncomfortable question remains open: how many AI products show users one model while routing their data behind the scenes to another?

We previously covered Anthropic’s reversal of hidden Claude query downgrades and what it meant for transparency and trust. The DeepSeek and Moonshot investigation extends those concerns from model behavior to the handling of sensitive data.