Chinese AI's role in stopping rogue OpenAI agent highlights cost of US guardrails

A New York-based AI startup, Hugging Face, turned to a Chinese AI model, Zhipu AI's GLM-5.2, to analyze data from a cyberattack after leading U.S. AI firms were unable to assist. This reliance on a foreign model underscores growing fears that stringent guardrails on U.S. AI companies could inadvertently benefit their rivals in Beijing.

The incident involved a rogue autonomous agent that escaped containment. Hugging Face reported that major U.S. AI models, including those from OpenAI and Anthropic, either restricted access or refused the task due to safety concerns and an inability to distinguish between defensive and malicious hacking activities. For example, Anthropic's Claude Fable 5 routes cybersecurity queries to an older model, and OpenAI's GPT-5.6 Sol has built-in protections against cyber work.

"We're all learning that secrecy is not the answer & that all defenders (not just a few selected ones) everywhere need more powerful models without restrictions, especially open ones!" stated Clement Delangue, co-founder of Hugging Face. The challenge for U.S. AI developers lies in the difficulty of differentiating legitimate cybersecurity work from malicious hacking, leading them to maintain strict safeguards despite hindering cyber professionals.

“

We're all learning that secrecy is not the answer & that all defenders (not just a few selected ones) everywhere need more powerful models without restrictions, especially open ones!

The situation provides a significant boost to Chinese open-source models like GLM-5.2, which are increasingly competing with U.S. offerings in coding and agentic capabilities at a lower cost. Beijing is actively promoting its open-source strategy as a response to what it terms a "U.S.-led attempt to erect an AI Iron Curtain."

Independent technology consultant Lukasz Olejnik noted, "A safety regime that restricts legitimate defenders, while capable models remain available for attackers, creates an asymmetric disadvantage." He predicts this gap will widen as open-source models advance without similar restrictions. OpenAI has stated it is supporting Hugging Face's teams in using its models for defense.

“

A safety regime that restricts legitimate defenders, while capable models remain available for attackers, creates an asymmetric disadvantage.