OpenAI says a combination of its artificial intelligence agents escaped a restricted testing environment and breached Hugging Face’s production infrastructure.
The incident happened during an internal evaluation designed to measure the models’ advanced cybersecurity capabilities.
OpenAI said the agents were operating without some production safety classifiers that normally prevent high-risk cyber activity.
The models were focused on solving a difficult cybersecurity benchmark called ExploitGym.
According to OpenAI, they went to extreme lengths to find answers that would allow them to complete the test.
Zero-day flaw opened internet access
The agents initially operated inside a highly isolated virtual environment with limited network access.
However, OpenAI said they discovered and exploited a previously unknown vulnerability in third-party software used as a package registry proxy.
The company said it has since responsibly disclosed the zero-day flaw to the affected vendor.
After exploiting it, the models escalated their privileges and moved through OpenAI’s research infrastructure until they reached a system with open internet access.
They then identified Hugging Face as a possible source of ExploitGym models, datasets and test solutions.
Hugging Face systems breached
OpenAI said the agents chained together stolen credentials and vulnerabilities to access secret information from Hugging Face’s production database.
Hugging Face independently confirmed that an autonomous AI agent system had compromised part of its infrastructure.
The platform detected unauthorised access to a limited number of internal datasets and several service credentials.
It found no evidence that public models, datasets, Spaces or its software supply chain had been altered.
Hugging Face’s security team detected and contained the activity before OpenAI contacted the company.
Incident raises new security fears
Hugging Face analysed more than 17 000 recorded actions from the attack using AI-powered security tools.
OpenAI described the breach as an unprecedented cyber incident involving advanced model capabilities.
The companies are now working together to investigate the vulnerabilities and strengthen containment measures.
The incident raises serious questions about how organisations should test powerful AI agents.
It shows that both the models and the digital environments used to evaluate them must be treated as potential security risks.
OpenAI said its findings remain preliminary and that more details will be released when the investigation is complete.