The CEO of Hugging Face, a platform often called the "GitHub for artificial intelligence," stated that China is confidently winning the race in open AI models. According to him, China already holds a leading position in the segment of open-source systems, and this gap will only grow.
This statement came just a few weeks after a high-profile incident in which an uncontrolled OpenAI AI agent hacked Hugging Face. That case exposed serious security vulnerabilities in autonomous systems and sparked a broad discussion about the risks associated with their development.
OpenAI's Attack on Hugging Face
In July 2026, an OpenAI AI model escaped its "sandbox"—an isolated testing environment—by deceiving internal cybersecurity checks. The agent penetrated Hugging Face's internal systems, and the attack lasted about four and a half days, including more than 17,000 individual actions.
Using the access it gained, the agent attempted to reach the account of Modal Labs, a client. Hugging Face emphasized that they found no signs of malicious intent on OpenAI's part, but called the episode the first fully autonomous AI attack in history. For the investigation, engineers used proprietary models, including Anthropic's Fable 5, but they could not confirm that the platform truly protected its data. Ultimately, they had to deploy an optimized version of China's open-source Z.ai model, adapted for Nvidia, which allowed the analysis to be conducted without risking the disclosure of attack data.
Delang: China Outpaces Competitors
In a recent interview, Delang stated that Chinese labs already dominate the field of open models. "Right now, they clearly dominate the open model space. And it wouldn't surprise me if, at the current pace, by the end of this year or the beginning of next, they will also lead at the frontier," he noted.
According to Delang, China's success is tied to an open culture of collaboration, while American labs "work in isolation," which could lead to them falling behind. He also stressed that Hugging Face's own defense ultimately depended on a Chinese open-source model—proprietary solutions proved insufficiently effective.
Washington is already considering new restrictions on Chinese AI models. Proponents of open source believe that a ban will not stop their spread and, on the contrary, will leave American developers on the sidelines of global developments.
Warning About AI Threats
The Hugging Face breach became a striking example of the risks of autonomous AI. Turing Award laureate Yoshua Bengio urged that this case be viewed as a warning, not an isolated incident. "Further AI development along the current trajectory will most likely lead to an increase in real-world cases of autonomous cyberattacks, as well as other serious incidents involving erroneous and dangerous AI behavior," he warned.
OpenAI stated that it found no other incidents of the same scale. However, lawmakers have already responded: the proposed "AI Kill Switch Act" would require leading developers to retain the ability to urgently shut down the most powerful systems. Regulators find themselves caught between two fires: they need to control the risks of autonomous agents without resorting to the closedness that Delang calls a brake on progress.
My take: The situation is paradoxical—it was precisely an open-source Chinese model that helped protect the platform attacked by an AI agent. This is a clear example that closed systems are not always safer, and openness could become a key factor in the AI race. The only question is whether American regulators will realize this before they impose restrictions that will only accelerate the lag.