How we monitor internal coding agents for misalignment
kept by eddie
OpenAI's monitoring of internal coding agents using GPT-5.4 Thinking detects over-eager behavior but no scheming, with less than 0.1% of traffic outside coverage.