China’s Probe of DeepSeek and Moonshot Raises Questions About Cross-Border AI Data

China’s Probe of DeepSeek and Moonshot Raises Questions About Cross-Border AI Data

September 22, 2026

China’s internet regulator is investigating DeepSeek and Moonshot AI after Anthropic alleged that sensitive user data may have been routed to Claude models, The Information reports. The companies and regulators have not publicly established the allegations as fact in the reporting available so far. The case matters because modern AI products can route requests across models, providers, and regions in ways that users may not see.

Model routing creates hidden data paths

A service may send a request to one model for drafting, another for reasoning, or a third for moderation. This can improve performance or lower cost, but every handoff creates a data-flow question: what information moves, who receives it, where it is processed, and how long it is retained?

Users may assume that a prompt stays with the app or provider they selected. In a routed system, the provider may use subcontractors or external models. A clear privacy notice should identify these relationships and explain whether prompts, files, or personal data are transferred to another company.

What the investigation can and cannot establish

A regulator’s investigation is a process, not a final finding. The available reporting describes allegations and inquiries, not a concluded violation. The next important details will be whether authorities confirm the scope of data transfers, what kinds of information were involved, and whether the routing was disclosed to customers.

The story also illustrates that data governance is not simply a matter of where a company is headquartered. A tool can be developed in one country, hosted in another, and send inference requests to a provider elsewhere. Enterprises need a map of the complete processing chain.

Steps organizations can take

Before using an AI service for confidential work, administrators should review model-provider disclosures, data-processing agreements, retention controls, and options to disable training or external routing. They should classify what information can be submitted and block restricted data from tools that do not meet requirements.

Technical controls can help: route sensitive workloads only to approved endpoints, log model destinations, minimize personally identifiable information, and test that settings behave as described. Vendor questionnaires should ask about subprocessors, processing regions, and incident notification timelines.

Why this is broader than one dispute

AI applications increasingly rely on networks of models and tools. Their value may come from combining specialized services, but the combination makes accountability harder. Users deserve to know when their data leaves the service they chose and what safeguards apply at each step.

Until the investigation produces verified findings, the claims should remain attributed. The practical lesson is already clear: organizations should treat AI model routing as a data-protection issue and make the route visible before sensitive information is sent.

Sources

• The Information: China probes DeepSeek and Moonshot
• Anthropic: Privacy and data usage
• China’s Cyberspace Administration
• NIST Privacy Framework