# China's Internet Regulator Opens Investigation Into DeepSeek and Moonshot Over Data Routed to Claude

Anthropic alleges that Moonshot sent more than 23 million exchanges to Claude between May and July, and that DeepSeek sent more than 12.1 million in a 14-day window. The allegations raise the question of whether Chinese user data reached United States (US) servers.

- Published: 2026-09-23T11:26:37.007Z
- Canonical: https://polylog.news/ai/2026-09-23/china-s-internet-regulator-opens-investigation-into-deepseek
- Publisher: Polylog (AI desk)
- Section: geopolitics
- Sources: [Polylog editors](https://polylog.news), [Anthropic News](https://www.anthropic.com/threat-intelligence-report-september-2026)

The Cyberspace Administration of China (CAC) has opened an investigation into DeepSeek and Moonshot AI over allegations that both companies routed user prompts to Anthropic's Claude models, [according to The Information](https://www.theinformation.com/articles/china-probes-deepseek-moonshot-potential-data-leaks-anthropic). The regulator first summoned representatives of all seven Chinese laboratories named in Anthropic's September threat report, then narrowed its focus to these two.

The underlying claims come from Anthropic's own [threat intelligence report](https://www.anthropic.com/threat-intelligence-report-september-2026). Anthropic alleges that Moonshot routed more than 23 million exchanges to Claude between May and July, and that DeepSeek routed more than 12.1 million over a 14-day period in July. It says the prompts contained names, email addresses, company data and other sensitive material belonging to hundreds of end users across at least a dozen languages, including material from multinational companies and state-affiliated actors. The CAC is reported to be examining whether police, military and state-linked corporate data reached US-operated models through this route. Neither DeepSeek nor Moonshot has publicly confirmed the routing, and Anthropic is a direct commercial competitor to both, a fact worth weighing given that only one party has the logs needed to prove the claim.

If substantiated, the allegation undermines a claim that has been central to the commercial case for Chinese open-weight models: that they form an independent technology stack, trained and served without dependence on American systems. A model provider that quietly relies on a US frontier API for difficult requests is selling capability it does not fully own. The episode also gives Washington a concrete argument for tightening access to US model endpoints, and gives Beijing a concrete argument for tightening data-export enforcement on its own companies. The reporting [comes ahead of scheduled Trump-Xi discussions](https://gizmodo.com/china-probes-deepseek-moonshot-ai-over-anthropics-claims-they-route-requests-to-claude-2000815507) that are expected to cover artificial intelligence.

## What this means

Enterprises outside the United States that adopted Chinese open-weight models specifically to keep data out of American jurisdiction now face a verification problem they cannot solve on their own: only the provider knows where a given request is actually processed. The likely result is demand for auditable inference, meaning attested serving locations and logged routing, which favors providers that host their own model weights over providers that resell an opaque endpoint. For Anthropic, publishing the allegation strengthens its enforcement position but invites the counterargument that it stands to benefit commercially from making the accusation.

## What to watch

- Whether DeepSeek or Moonshot publishes routing logs or a technical rebuttal. Silence would likely be read by enterprise buyers in both China and the West as confirmation.
- Whether the Cyberspace Administration of China issues a data-export penalty, which would set the first concrete precedent for punishing a domestic laboratory over dependence on a foreign model.
- Whether other model resellers begin publishing attested serving-location guarantees in their contracts.