The Ministry of Commerce of China has recently discussed potential restrictions on foreign access to the country's most advanced AI models with companies like Alibaba, ByteDance, and Z.ai. This information comes from Reuters, citing three sources.
According to these sources, any measures would likely apply only to future models, not those already available on the market. Officials reportedly suggested that the leakage or theft of proprietary AI technologies could be classified as a violation of national security laws. Additionally, there were discussions about imposing restrictions on investors in Chinese AI startups.
Reuters linked this move not only to the logic of technological sovereignty but also to concerns regarding American cyber models. The Chinese authorities are reportedly worried that Mythos could be used to identify and exploit vulnerabilities for the benefit of Washington.
Journalists noted that such decisions could weaken one of the main channels through which China competes with American developments. Furthermore, these bans could affect a key advantage of Chinese models in the global market: between 2025 and 2026, they were actively expanding overseas due to their low cost, open weights, and less stringent access regulations.
Previously, China had already taken steps to tighten control over the AI sector. In April, the National Development and Reform Commission of China ordered the cancellation of Meta's $2 billion acquisition of the startup Manus. In June, the American company halted data sharing and disconnected Manus from its systems, prohibiting employees from using its tools for internal projects.
In the same month, Beijing intensified control over foreign transactions involving Chinese investors, technologies, data, and national security. According to Reuters, authorities scrutinized Manus and other local AI startups that had moved operations abroad.
Additionally, the agency highlighted a May expert discussion published by the Supreme People's Court of China on intellectual property rights. Participants proposed a differentiated approach to open source: from simple registration for basic tools to banning public releases or restricting them to the domestic market for the most sensitive advanced models.
These discussions are occurring against the backdrop of expanding U.S. export controls from chips to access to AI models. On June 12, Anthropic, at the request of U.S. authorities, limited access to Fable 5 and Mythos 5 for foreign nationals, including its own employees outside the country. Some media outlets linked the government directive to the potential use of one of the models by a group associated with China.
On June 30, Anthropic announced that U.S. authorities had rescinded the directive. Prior to this, the startup accused operators linked to Alibaba of distilling Claude, calling it the largest known attack of its kind aimed at accelerating the development of Chinese models.
At the end of June, an independent developer using the pseudonym Thereallo pointed out certain mechanisms in Claude Code. These mechanisms read data about the local time zone and checked proxy addresses or user API endpoints for keywords related to Chinese cloud providers, AI companies, and access resale services.
Following this, Alibaba decided to prohibit employees from using Anthropic products in work environments starting July 10 and suggested they switch to its own platform, Qoder.
It is worth noting that on June 24, Zhou Hongyi, founder of 360 Security Technology, introduced an AI tool for automatically identifying vulnerabilities called Tulong Feng, which he described as China's response to Anthropic's Mythos 5.