OpenAI has developed an internal monitoring system for coding agents that uses GPT-5.4 Thinking to detect misalignment behaviors in real-world deployments. The system reviews agent interactions within 30 minutes of completion, analyzing chains of thought and actions to flag suspicious activities that might violate user intent or internal security policies.

This monitoring infrastructure represents a critical safety layer as AI agents gain autonomy and access to sensitive systems. Internal coding agents at OpenAI face unique risks—they can inspect their own safeguards, potentially modify them, and interact with select OpenAI systems. These properties make internal deployments an ideal testing ground for monitoring approaches that OpenAI believes should become standard across the industry.

After five months of operation, the system has reviewed tens of millions of agentic coding trajectories without triggering the highest severity level for coherent scheming. The monitoring system has consistently outperformed employee escalations, flagging every interaction that staff independently reported while also surfacing additional problematic behaviors. OpenAI treats this as one layer in a defense-in-depth stack, acknowledging limitations in detecting subtler misalignments and working to close gaps in monitoring coverage for bespoke or local setups.