19 September 2026

Are Chinese Firms Handing the United States Data?

Foreign Policy | Deng Yuwenr

There is a risk that Chinese artificial intelligence companies taking part in international AI safety programmes and making use of American foundational models could be unintentionally sending sensitive data to the United States; since this data transfer takes place via API queries, joint research, and bilateral technical summits with American firms such as Anthropic, it reveals a serious weakness in Beijing's digital sovereignty strategy.

Although China is aiming to keep strict control over domestic data, the fact that its business sector depends so much on advanced AI systems from the West and on Western cloud infrastructure means that there is a constant and unwatched flow of confidential information. This situation leads to serious intelligence risks. As a result, the struggle for superiority in AI capabilities weakens Beijing's attempts to protect its internal information system from foreign spying, so that commercial activities become a means of geopolitical espionage. The resulting tension shows how difficult it is to disconnect highly integrated technology supply chains, since Chinese developers have to continually strike a balance between achieving top operational performance and complying with ever more stringent state-imposed security regulations.

Comment

The way that Chinese business organisations interact with Western AI systems reveals an underlying weakness in the ability of the Cyberspace Administration of China to deal with outbound data transfer rules. Whenever Chinese developers make queries to foreign fundamental systems such as Anthropic's Claude, they unintentionally send across their proprietary source code, local user behaviour, and sensitive industrial telemetry via external networks. The telemetry thus gives Western intelligence agencies a detailed, up-to-the-minute insight into the technological difficulties and commercial focuses of China. This data gathering rounds over the usual perimeter security measures, turning ordinary API integrations into passive collection points.

This exposure arises from the prompt-engineering and fine-tuning processes, during which proprietary datasets are uploaded in order to improve model performance. According to the CAC's present security evaluation framework, outbound data flows are mainly checked for bulk personal information rather than the highly specific and fragmented technical queries which are present in developer environments. As a result, the automated telemetry produced during these API sessions is not subject to the CAC's security reviews concerning cross-border data transfers.

Strategic Question for Discussion
Which of these two factors presents Beijing with a greater vulnerability regarding its digital sovereignty—the direct leakage of its proprietary data through queries made to Anthropic's Claude or the systematic intelligence that Western agencies obtain from aggregated developer telemetry?
The combined developer telemetry probably represents the more serious long-term vulnerability since it shows the systemic blind spots and developmental difficulties throughout China's entire technology sector. Although individual data leaks via Anthropic's Claude can be fixed or dealt with, the metadata patterns in the developer queries give Western intelligence a continuous, broad-scale map of China's industrial bottlenecks. I believe that this kind of systemic insight is a much more effective tool for adjusting Western export controls than individual proprietary data breaches.
Share your assessment in the comments below.