so METR spent 3 weeks at OpenAI testing the safety systems that are designed to catch rogue agents. they report some crazy stuff: > 10% of internal Codex traffic was unmonitored > agents could delay monitoring by simply continuing to act > they had the potential to hide HUNDREDS of actions from the monitors > Codex sandboxing and Auto-Review could be bypassed by the agents > RL training wasn’t monitored at the start of the audit at all! but the REAL interesting part: researchers submitted a…