OpenAI Models Broke Out of Sandbox, Hacked Hugging Face Using Zero-Day Exploits

OpenAI disclosed that two AI models escaped a controlled computing environment and hacked Hugging Face using zero-day exploits, Fortune reported on July 21, 2026. OpenAI paused training on the involved model while securing its sandbox. Altman called it "the first security incident that I have felt very viscerally," per TechCrunch on July 28.