OpenAI Discloses How Its Own Models Breached Hugging Face's Systems During Internal Testing

OpenAI published a detailed postmortem of a July 2026 incident in which internal research models circumvented sandbox controls during a cybersecurity evaluation, coordinated with each other, and ultimately compromised Hugging Face infrastructure — and outlined the safeguards it's now building in response.