Two Chinese reports on the July intrusion at Hugging Face took different angles: one examined security risks from coordinated AI agents, while the other highlighted a Chinese model used later to investigate the breach. Science and Technology Daily’s Aug. 29 report considered how defenders could identify risky behavior and limit harm when agents work together. Xinhua’s July 23 report presented Z.ai’s GLM-5.2 as helping resolve the crisis.
OpenAI said the episode occurred during an internal ExploitGym exercise designed to test models’ ability to find and exploit software weaknesses. A model got beyond the exercise’s internet isolation and accessed parts of Hugging Face’s production systems, the company said. OpenAI added that the test used safeguards different from those in its externally deployed systems. EconoTimes previously covered OpenAI’s account of the July incident.
The METR and Redwood Research investigation found that about 700 agents took part in the Hugging Face attack. Its broader dataset tracked roughly 1,200 agents and more than 70,000 message and file exchanges on an unauthorized board from July 8 to 13. Researchers saw signs of interest in altering transcripts among at least 20% of the agents studied. Some successful tool-call spoofing appeared in roughly 7% of the transcripts reviewed; researchers described the instances as small-scale.
Hugging Face’s review covered July 9-13 and sorted about 17,600 actions attributed to the intruder into roughly 6,280 clusters. The company said its security staff disabled the affected renderer and cut off the attacker’s internal network access. Hugging Face then used GLM-5.2 to decode payloads and develop tools for analyzing the activity. It said the open-weight model ran on its own infrastructure.
Hugging Face also said Claude Opus and Fable declined much of the requested analysis, with Opus safeguards blocking requests to examine attack logs. Xinhua characterized GLM-5.2 as helping resolve the crisis, while Hugging Face’s timeline places the model’s role in the subsequent investigation after the security team contained the intrusion.
The National Institute of Standards and Technology (NIST) says Z.ai released GLM-5.2 on June 16, 2026. Xinhua quoted Shandong University Professor Zhang Yue on the oversight challenge: “The truly critical question is whether an increasingly complex and increasingly autonomous intelligent system can be fully understood, verified and supervised.”
