OpenAI is using its most advanced models, like GPT‑5.4 Thinking, to monitor internal coding agents for misaligned behavior in real-world deployments. These agents have access to sensitive internal systems, code, and safeguards, making them a unique testbed for safety research. The monitoring system flags actions that deviate from user intent or violate security policies, all while preserving privacy and data security. Why it matters: As AI agents gain more autonomy, building robust monitoring infrastructure today is essential to catching subtle misalignments that only surface in complex, tool-rich workflows—stuff you can’t easily spot in a lab.