OpenAI unveils new RL training safety protocols

OpenAI has introduced new safety protocols targeting frontier reinforcement-learning training, covering alignment, containment, and monitoring controls. Plans include automated tools and human reviewers to identify weak spots in RL datasets, multi-layer infrastructure security, and repeated red-teaming of sandboxes. NVIDIA also announced its Open Agent Safety Platform on Monday.