OpenAI Reveals Its Layered System for Catching Misaligned AI Agents
OpenAI says replayed monitors would have flagged the Hugging Face breach 30 hours early. Its L0–L3 system pages humans in 15 minutes across RL, evals, Codex, and external deployment.