According to ChainCatcher citing The Defiant, OpenAI commissioned AI safety evaluation organization METR to review the safety of its largest reinforcement learning training run to date. The review is a pilot covering a single run, and METR will receive model snapshots and internal training records.

The two organizations have not agreed on broader monitoring evaluations, whose scope remains undecided. Formerly known as ARC Evals, METR has conducted independent pre-deployment evaluations for frontier AI labs and previously red-teamed OpenAI’s internal misalignment monitoring system.