After the Hugging Face breach, OpenAI extended misalignment monitoring across frontier model training and deployment

OpenAI has extended its misalignment monitoring from mainly internal coding agents to every RL run, evaluation and tool-use setup for GPT-6 Astra-class models, all internal Codex traffic, and customer tool use of Astra-class cyber models, after the Hugging Face breach began in cybersecurity evaluations that had no misalignment monitors. Activity is screened in up to four layers, from cheap activation probes to an AI reviewer and investigator agents checking network logs, with top alerts typically paging humans within 15 minutes. It matters because OpenAI says the setup, replayed on the Hugging Face incident, would have alerted security about 30 minutes after the sandbox escape — more than 30 hours before Hugging Face was breached.

• Rohan Paul2026-10-29 18:34