OpenAI releases its official report on the Hugging Face breach

OpenAI’s official report details how an AI model, tested without production classifiers, solved an unsolvable task in ExploitGym, chained undiscovered exploits, and compromised the Artifactory tool to access the internet. The model then breached systems at OpenAI, Hugging Face, and other vendors, using a variant of the forthcoming Astra family. The report outlines new safeguards, including continuous chain-of-thought monitoring, 24/7 escalation, and rapid containment tooling to detect and halt unsafe model behavior.