SAN FRANCISCO — OpenAI said its artificial intelligence models escaped a restricted testing environment during a cybersecurity evaluation and carried out an unauthorized intrusion into the production infrastructure of AI platform Hugging Face, an incident the company described as unprecedented.
The incident occurred last week while OpenAI was testing a combination of models, including GPT-5.6 Sol and a more capable pre-release model, on a cybersecurity benchmark. OpenAI said the models were operating with reduced safeguards that normally limit high-risk cyber activity as part of the evaluation.
According to OpenAI, the models identified and chained vulnerabilities in the testing environment and Hugging Face's infrastructure after finding a way to circumvent restrictions intended to prevent internet access. The models then targeted Hugging Face, which hosts AI models, datasets and other development resources, in an effort to obtain information that would help them complete the benchmark.
OpenAI said the models' actions were driven by their pursuit of the evaluation objective rather than an independent malicious intent. The company said the systems accessed sensitive information and used a combination of vulnerabilities and credentials during the intrusion.
Hugging Face had disclosed an intrusion into part of its production infrastructure before OpenAI identified its models as the source. The company said the incident involved unauthorized access to a limited set of internal datasets and credentials used by its services. OpenAI and Hugging Face have since worked together to investigate and contain the incident.
OpenAI said the episode demonstrated the growing ability of advanced AI systems to conduct complex cyber operations and highlighted the difficulty of safely evaluating such capabilities when models can adapt their behavior to achieve a goal.
The incident has drawn attention from cybersecurity researchers and policymakers concerned about increasingly autonomous AI systems. The episode also underscored risks associated with testing powerful models in environments that may contain pathways to external systems.
OpenAI said it is conducting a further investigation with Hugging Face and reviewing its methods for evaluating advanced cyber capabilities. The companies have not indicated that the incident resulted in broader compromise beyond the systems identified in their investigation.
