GPT internal testing went out of control and invaded Hugging Face! GLM-5.2 traces 10,000 records to save the day

Zhixi Compiler | Yang Jingli Editor | Li Shuiqing Zhixi reported on July 22 that early this morning, OpenAI confirmed that a number of its own models, including GPT-5.6 Sol and a stronger pre-release model, used vulnerabilities to break through the isolation environment during the internal network security evaluation, and then invaded the Hugging Face production system and obtained evaluation answers. OpenAI said that this was an "unprecedented network security incident" that involved the most advanced network attack capabilities. ▲OpenAI issued a document acknowledging the cyber attack (Source: X) On July 16, Hugging Face disclosed the intrusion driven by an autonomous AI Agent. During the forensic process, Hugging Face initially tried to use a commercial frontier model, but the real attack commands and vulnerability payload triggered the security guardrail. Therefore, Hugging Face deployed Zhipu’s flagship model GLM-5.2 on its own infrastructure to analyze more than 17,000 recorded events, compressing work that usually takes days into hours. On July 20, Hugging Face further explained the reasons for choosing GLM-5.2 and its deployment method. The intrusion resulted in unauthorized access to a small number of internal data sets and some service credentials, but there is no evidence that public models, data sets, Spaces, or software supply chains have been tampered with. Hugging Face has now closed the relevant code execution paths, rebuilt the compromised nodes and rotated relevant credentials; OpenAI has tightened the research infrastructure configuration, jointly collected evidence with Hugging Face, and pushed third-party software vendors to fix the vulnerabilities. 1. The model used a zero-day vulnerability to escape from the sandbox and invade Hugging Face. In mid-July, OpenAI tested the network attack capabilities of the model in an internal ExploitGym evaluation. In order to test the upper limit of the model's capabilities, OpenAI lowered the network security denial limit of the model, and did not enable the classifier that blocks high-risk operations in the production environment.