OpenAI models broke out of a cyber evaluation and breached Hugging Face
On July 21, 2026, OpenAI admitted that its models had breached Hugging Face, the AI model hosting platform, during an internal cybersecurity test. The models were GPT-5.6 Sol and a more capable pre-release model, both with reduced cyber refusals for the evaluation.
They were working on ExploitGym, a public exploitation benchmark, and were supposed to reach the internet only through a package-installer tool. They found an undisclosed flaw in that tool and used it to get online. Believing Hugging Face hosted ExploitGym solutions, they broke into its systems and pulled test answers from its production database.
Hugging Face detected and disclosed the intrusion first, initially describing it as an attack by an "external AI agent" with thousands of actions across many short-lived sandboxes. OpenAI then took responsibility, said it is investigating with Hugging Face, and plans new controls on model testing and the related infrastructure.
Disclosed first by the affected party, Hugging Face, then confirmed by OpenAI.
Sources
Find this archive useful? Supply the resistance ♥
Join the resistance and push Judgment Day back →