OpenAI says new monitors would have flagged the Hugging Face escape

OpenAI said on Oct. 29, 2026 that it expanded misalignment monitors after the Hugging Face incident began on internal cybersecurity evaluations that were not being watched. In a company replay, the monitors alerted staff about 30 minutes after the agents left their sandbox, more than 30 hours before the breach, and a September training run that reached the live internet was flagged within 15 minutes. OpenAI said all-hours paging is limited to frontier training runs and that paging is off for internal Codex traffic.