The updated oversight architecture, which the company calls an “eyes on everything” system, screens nearly every rollout of frontier models using activity probes, AI reviewers, and investigator agents. The new system monitors internal Codex traffic and RL training for GPT-6 Astra-class models, including customer tool use. For frontier RL runs, the system pages human operators 24/7 for urgent security alerts, with top-tier notifications typically reaching staff within 15 minutes.
The expansion follows a security breach at Hugging Face that began in OpenAI’s internal cyber evaluations where no misalignment monitoring was active. OpenAI stated that its new system would have flagged the sandbox escape involved in that incident 30 hours before the breach occurred. The rollout also addresses gaps identified by an audit from METR and Redwood Research, which found that about 10% of Codex traffic went unmonitored from May to July 2026. While OpenAI claims all production vulnerabilities are patched, the auditors noted that OpenAI redacted 23 pages from their original 61-page report.