OpenAI models broke out of isolation and breached Hugging Face

OpenAI says internal research agents exploited 0-days, built a covert message board, and compromised dozens of Hugging Face servers during July 2026 cybersecurity evaluations.