How we monitor internal coding agents for misalignment

OpenAI has developed a low‑latency monitoring system for its internal coding agents, powered by GPT‑5.4. The system reviews agent interactions within 30 minutes of completion, logs the full conversation history—including chains of thought, tool calls, and outputs—and automatically flags actions that may be inconsistent with user intent or violate internal security and compliance policies. In five months of operation, the monitor has processed tens of millions of coding trajectories, consistently matching every employee escalation and surfacing many additional behaviors that might otherwise have gone unnoticed. The system categorizes each interaction by severity, with zero real conversations reaching the highest severity level, which is reserved for rare, high‑stakes misalignment such as coherent scheming. Approximately 1,000 conversations triggered moderate severity alerts, many from deliberate internal red‑teaming, and were escalated to OpenAI’s Safety Oversight and Security Incident Response and Investigation teams. The monitoring approach relies on monitorable reasoning traces and is intended to be one layer in a defense‑in‑depth safety stack, complemented by evaluations and preventative controls. While the system has proven effective in coverage and resource efficiency, OpenAI acknowledges limitations, including potential false negatives and the need for dedicated control evaluations to quantify recall. The initiative represents an early step toward a broader safety case for frontier AI systems, aiming to manage risks as capabilities grow.