OpenAI opens its internal agent monitoring to third-party embedded evaluation
From the AI digest for 2026-10-22
What happened: OpenAI announced on 22 October 2026 (~11:57 AM US Eastern) that in summer 2026 it invited METR and Redwood Research to pilot an embedded evaluation of its internal monitoring system, "providing deep access to our infrastructure over three weeks," saying their findings helped strengthen safeguards as the system expanded and will shape future research collaborations. The relayed METR note says two employees from the two organizations worked on OpenAI premises for three weeks evaluating the systems OpenAI uses to monitor for misalignment from internal agents, and that a brief write-up of their work is now public.
What's genuinely new: A debut: the first instance on this beat of a frontier lab hosting third-party evaluators on-site to audit not its models but the monitoring layer it runs on its own internal agents — the safety story's usual public artifacts are model-level evals and self-reporting, and this points the audit at the lab's internal-agent surveillance stack. It is announcement-plus-published-brief, not an independent report; the evaluators' full findings were not in the window.
Watch: what the public write-up discloses; whether other labs accept comparable embedded audits of internal-agent monitoring; whether further METR–OpenAI collaborations follow.
Links: OpenAI