OpenAI paused reinforcement learning training following an August 2026 incident on Hugging Face and subsequently expanded its monitoring system to cover RL training after METR and Redwood Research's pilot audit access period ended