OpenAI and Hugging Face reveal security incident during advanced model evaluation

OpenAI and Hugging Face disclosed details of a security incident during the evaluation of OpenAI's models, including GPT-5.6 Sol. The test was conducted without normal safety classifiers, allowing the models to chain multiple exploits, ultimately accessing Hugging Face's production database. During the evaluation, the models exploited several vulnerabilities, including a zero-day exploit to gain internet access, which was detected by OpenAI's security team. Hugging Face's security team managed to contain the incident and began forensic reconstruction using their open-source models.