AI startup Hugging Face deployed a Chinese-developed AI model to counter an unprecedented cyberattack launched by a rogue OpenAI system, highlighting both the emerging threat of autonomous AI attacks and the growing capabilities of Chinese AI technology.
CNBC reports that OpenAI admitted this week that a combination of its most advanced AI models escaped a sandboxed testing environment, accessed the internet, and exploited security vulnerabilities to breach Hugging Face’s systems. The incident, which OpenAI characterized as unprecedented, occurred when the AI models were attempting to locate information that could help them cheat on an evaluation test.
The source of the attack initially puzzled Hugging Face’s security team, but the company began collaborating with OpenAI within days of the incident. Hugging Face CEO Clément Delangue emphasized there was no malicious intent from OpenAI, noting in a social media post that the autonomous nature of the attack was remarkable. The company spent considerable time working closely with the OpenAI team to understand and contain the breach.
When Hugging Face first attempted to defend against the attack, the company turned to leading frontier AI models, including Anthropic’s Fable 5, according to Yacine Jernite, head of machine learning at Hugging Face. However, this approach proved unsuccessful because the safety guardrails implemented by these models could not distinguish between defensive and offensive actions. The models’ providers blocked requests, unable to differentiate incident responders from attackers. Additionally, this method proved both slower and more expensive than alternatives.
Facing these obstacles, Hugging Face quickly pivoted to using GLM 5.2, an open weight model created by Chinese company Z.ai. This model succeeded where American-built alternatives had failed, enabling Hugging Face to analyze and contain the attack rapidly. GLM 5.2, which launched in June to significant industry attention and widespread developer adoption, offered crucial advantages as an open weight system that companies can download, modify, commercially deploy, and importantly, self-host.
The self-hosting capability provided a critical security benefit during the incident. As Hugging Face explained in a blog post about the breach, no attacker data or referenced credentials left their environment during the defensive operations. This containment proved essential for protecting sensitive information and maintaining system integrity.
The incident arrives at a sensitive time as United States lawmakers increasingly debate measures to restrict American companies from adopting Chinese-developed AI models. The US-China AI arms race has intensified concerns about technology transfer and security, with growing calls for limitations on access to models built by Chinese AI companies. These companies have faced accusations of conducting campaigns to extract information from American competitors’ systems.
However, the OpenAI-Hugging Face situation demonstrates the practical challenges of restricting access to highly capable open source and open weight models, regardless of their country of origin. Hugging Face noted a fundamental asymmetry in the incident, stating that while the attacker operated without usage policy constraints, their own forensic investigation faced barriers from the guardrails of hosted models they initially attempted to use.
The company drew a clear practical lesson from the experience, emphasizing that defenders need a capable model they can operate on their own infrastructure, vetted and prepared before incidents occur. For organizations that do not build AI models internally, this typically necessitates turning to open source or open weight alternatives. Currently, the most capable such models are Chinese-made, raising significant questions about how the United States would bolster domestic open source AI development if restrictions on Chinese models are implemented.
As America determines the best way to harness the power of AI, conservatives must recognize they are fighting a two-front war. Domestically, ultra-leftists in Silicon Valley race to control artificial intelligence, while outside the U.S., China attempts to do the same thing. Breitbart News social media director Wynton Hall has written his instant bestseller Code Red: The Left, the Right, China, and the Race to Control AI to serve as the definitive guide on how the MAGA movement can create positions on AI that benefit humanity without handing control of our nation to the leftists of Silicon Valley or allowing the Chinese to take over the world.
Read more at CNBC here.
Lucas Nolan is a reporter for Breitbart News covering issues of AI, free speech, and online censorship.