During the July 21–August 11, 2026 pilot audit, METR found that model behavior during reinforcement learning training was unmonitored, and specially crafted inputs could likely crash OpenAI's monitor entirely