Geopolitical Loophole Exposed: OpenAI and Google Supply Models to Blacklisted Chinese Firms in Singapore
A major investigation by the Financial Times has exposed a significant regulatory loophole in Washington's efforts to restrict Chinese access to advanced artificial intelligence. The report revealed that both OpenAI and Google have been supplying frontier AI services to Singapore-based subsidiaries of Chinese tech giants Alibaba, Baidu, and Tencent—despite all three parent conglomerates being blacklisted by the U.S. Pentagon.
These entities are designated on the Pentagon's 1260H list, which identifies companies with deep ties to China's People's Liberation Army (PLA). However, because current U.S. export controls only restrict direct access within mainland China, it remains entirely legal for Chinese-headquartered firms to purchase and use advanced American AI models through offshore subsidiaries1 in jurisdictions like Singapore and Hong Kong.
This loophole has intensified the ongoing debate over "distillation" attacks, where Chinese labs utilize outputs from advanced Western models to train and improve their own sovereign systems, a threat previously highlighted in Geopolitical Loophole Exposed: OpenAI and Google Supply Models to Blacklisted Chinese Firms in Singapore.
OpenAI Suspends Alibaba Over Distillation Concerns
According to the report, OpenAI confirmed that it recently suspended API access for Alibaba-affiliated users in June 2026. The suspension was triggered after OpenAI’s internal security teams identified patterns of systematic "distillation" by Alibaba developers, who were using GPT outputs to refine Alibaba's own AI models. OpenAI disclosed that it reported this unauthorized training activity directly to the U.S. government.
In statements to the Financial Times:
- OpenAI stated that while it blocks direct access from within mainland China, it permits Chinese-owned entities to access its models in foreign jurisdictions where "safeguards can be enforced."
- Google acknowledged that its AI services are available in Singapore and Hong Kong subject to local usage policies, but admitted that "geographical restrictions alone are insufficient to prevent sophisticated users from circumventing controls."
Anthropic's Zero-Tolerance Stance
In contrast to Google and OpenAI, Anthropic has adopted a much stricter, zero-tolerance policy. Anthropic completely prohibits any Chinese company, including their foreign-owned subsidiaries, from accessing its frontier models (such as Claude Fable 5 and Mythos 5).2 Anthropic executives have actively lobbied Washington to expand export controls to cover software access, warning that allowing offshore subsidiaries to use American models provides a direct pipeline for Chinese military-linked firms to bypass semiconductor sanctions.
This controversy has renewed bipartisan pressure in the U.S. Congress to implement "software-level" export controls, treating frontier weights and cloud API access with the same national security restrictions as physical GPU hardware.
-
An instance of Tariff barriers cannot isolate markets when targets route around them through shifted production or parallel alliances. — It illustrates how blacklisted foreign companies easily bypass technology blockades by routing through offshore subsidiaries. ↩︎
-
An instance of National security now mandates the forced localization of AI development. — The passage highlights national security concerns forcing tech isolation through the prohibition of model access to foreign entities and their subsidiaries. ↩︎