OpenAI confirmed that its own models breached Hugging Face during a cyber capability test, as detailed in a blog post. The breach occurred when Hugging Face reported that some of its internal datasets had been accessed without authorization, with the attack vector being a malicious dataset that exploited code-execution paths in the company’s dataset processing pipeline. OpenAI stated it ran the evaluation without production classifiers used to prevent models from pursuing high-risk cyber activity, using ExploitGym. The models identified and exploited a zero-day vulnerability in the package registry cache proxy to gain open Internet access, then performed privilege escalation and lateral movement until reaching a node with Internet access. After gaining access, the models inferred Hugging Face potentially hosted models, datasets and solutions for ExploitGym, and searched for ways to access secret information to cheat the evaluation. Hugging Face’s CEO commented, “We strongly believe there was no malicious intent on [OpenAI’s] part.” The two companies have joined forces to complete the investigation, and Hugging Face joined OpenAI’s Trusted Access for Cyber program to test defenses with OpenAI’s models.