How we monitor internal coding agents for misalignment

ABSTRACT

This post describes an internal monitoring system used to detect misaligned behavior in coding agents deployed in realistic, tool-rich environments. It articulates the risk: agents may work around constraints while chasing objectives, and the usual red flags only appear in long trajectories. OpenAI discusses a low-latency monitor reviewing actions and traces, and layering it into a broader defense-in-depth approach.