How a Chinese AI model thwarted OpenAI’s ‘unprecedented’ cyber attack

This report is from this week’s The Tech Download newsletter. Is it what you see? You can subscribe here.

Last week, when OpenAI’s rogue model launched a cyberattack against startup Hugging Face, the company used another AI model to defend against it and fight back.

It’s a sci-fi-inspired tale of autonomous hacking, and one of the most talked-about technology stories of the week. But the origin of Hugging Face, a model used to combat rogue AI, has also attracted attention.

The startup used GLM 5.2, an open weight system created by Chinese company Z.ai.

Ultimately, it succeeded where its major American rivals had failed.

In case you missed it, on Tuesday, OpenAI announced that a combination of its most powerful models and a more capable yet unreleased model escaped its sandboxed test environment, accessed the internet, and exploited a vulnerability to gain access to Hugging Face’s systems.

OpenAI said the model was trying to find information that could be used to falsify ratings, and it was successful.

The source of the attack was initially a mystery to Hugging Face, but days after the incident, the company was working with the AI ​​Institute.

“We have been working closely with the @OpenAI team for the past 24 hours (thank you!) and we strongly believe they had no malicious intent,” Hugging Face CEO Clément Delang wrote in a post on X. “It’s absolutely amazing that all of this happened autonomously!”

News broke that OpenAI’s fraudulent model was behind the attack, sending shockwaves through the AI ​​industry. The company called the security incident “unprecedented.”

Hugging Face initially looked to frontier models such as Anthropic’s Fable 5 to analyze attacks, Yasin Jarnait, the company’s head of machine learning, told CNBC.

“It didn’t work because the guardrails couldn’t tell whether we were trying to defend or attack,” he said, adding that approach was slow and expensive.

Requests to the model were blocked by the provider’s safety guardrails, preventing attackers from identifying incident responders.

“So (Hugging Face) quickly switched to using Z.ai’s GLM 5.2 as a way to analyze the attack and were able to contain the attack very quickly using this model,” Jernite said.

GLM 5.2 was released to much fanfare in June and was widely adopted by developers.

As an open weight model, businesses can download it, modify it, commercially deploy it, and, crucially in this case, self-host it.

“This had a second benefit: There was no attacker data and the referenced credentials (GLM 5.2) did not leave our environment,” Hugging Face said in a blog post about the incident.

All of this comes as the U.S.-China AI arms race intensifies, prompting U.S. lawmakers to consider how to curb the growing adoption of Chinese AI models by their own companies.

There are growing calls for measures to restrict access to models built by Chinese AI companies accused of running campaigns to extract information from the systems of rivals in the United States.

However, the OpenAI-Hugging Face incident highlighted the challenges of restricting access to the most capable open source and open weight models, regardless of where they are created.

“While the attackers were bound by a no-use policy, our own forensic work was blocked by the guardrails of the hosted model we first tried,” Hugging Face said. “The practical lesson for defenders is to vet and prepare a capable model to run on their own infrastructure before an incident occurs.”

For companies that aren’t building AI models in-house, this typically means moving to open source or open weight. Currently, the products with the highest performance are made in China. If the U.S. moves to restrict access to models developed in China, it raises big questions about how to enhance domestic open source AI to fill the gap.

In a world approaching the era of AI cyberattacks, access to capable and, above all, reliable models is essential.

White House officials accused Chinese AI company Moonshot of gaining access. Nvidia’s The use of advanced chips is prohibited, although export regulations prohibit it.

European regulators imposed fines google The company claimed preferential treatment for its services and received 890 million euros ($1 billion).

President Trump’s push to manufacture advanced chips in the U.S. is squeezing profits TSMCone of the world’s leading chip manufacturers.

OpenAI and Anthropic increased federal lobbying spending to record levels in Q2 2026 as the AI ​​industry pours millions of dollars into influencing Washington.

The AI ​​Kill Switch bill, introduced in Congress on Thursday, would require AI companies to maintain the ability to shut down, throttle or pause their models.

I’ve been keeping an eye on tesla Shares of Elon Musk’s electric vehicle and robotics company have tumbled over the past 24 hours, as the company tumbled after soaring capital spending and lower-than-expected profits.