The @OpenAI incident report on the @HuggingFace breach explains how agents escaped a cyber test and why its safeguards failed. OpenAI says its production setup cuts the behavior by over 100×. It is also isolating sandboxes, restricting internet access and expanding monitoring. (Digg)