As more Chinese developers covertly use Anthropic's Claude and OpenAI's GPT, the Chinese government has issued another security warning. The warning targets "AI relay services," channels used to bypass restrictions and access advanced AI blocked by the United States. About 10 days after the U.S. specifically targeted China's "covert use," the Chinese government has joined in restricting such use, though for a different reason.
According to the South China Morning Post (SCMP) on Tuesday, China's Ministry of State Security (MSS) posted a warning on its official WeChat account on Monday. It said that using AI relay services, which help access overseas AI models, carries risks of data leaks, privacy violations, and unauthorized cross-border data transfers.
AI relay services are intermediary platforms that connect local developers with AI providers. Through a single interface, they allow simultaneous access to multiple domestic and overseas models. Their strengths include convenience, low cost, and access to models unavailable within China. However, authorities pointed out that some operators heighten security risks.
The MSS said that "some platforms operate without proper qualifications and with weak security controls," and that they "increase the risk of privacy leaks and illegal data trading."
Neither Claude nor GPT supports users in mainland China. As a result, more Chinese developers have turned to relay services as a workaround.
Earlier, the U.S. announced guidelines on advanced AI semiconductors on the 31st of last month. Under the rules, companies headquartered in mainland China or Macau must obtain prior license approval from the U.S. government when importing or exporting advanced American computing products, such as those from Nvidia, even through third-country entities.
As a result, with direct imports of Nvidia's top-tier chips blocked, Chinese big tech firms and AI startups turned overseas. They have procured the chip-based computing power needed to train next-generation large language models (LLMs) from overseas data centers in Southeast Asia and elsewhere. Alibaba and ByteDance are also known to have accessed key Nvidia chips by concealing their nationality through local data centers in Southeast Asia.
At the time, China's Ministry of Commerce strongly criticized the U.S. guidelines, saying they "abuse export controls and seriously disrupt the global semiconductor supply chain."
Now, that same ministry has done the opposite, pointing to its own tech companies' workaround use of U.S. AI models. It said some platforms store user data without encryption, creating a risk of sensitive information leaks. It also noted that they could deceive customers by swapping advanced models for cheaper alternatives.
The risk of "backdoors" was also raised. A backdoor is a hidden method of accessing a system by bypassing authentication and security controls. If exploited, it can be used to hijack accounts or surveil developers' devices.
Meanwhile, the U.S. warned in April that China was attempting large-scale access to advanced American AI systems through networks of intermediary accounts. Anthropic had disclosed attempts by China-linked actors to access its models through organized proxy infrastructure (networks of relay servers that ensure anonymity).
