OpenAI builds GPT-5.4-powered monitor for its internal coding agents
OpenAI has built a low-latency monitoring system, powered by GPT-5.4 Thinking at maximum reasoning effort, to detect misaligned behavior in the coding agents it deploys internally. The company described the system in a March 19 post, saying it was built and refined over the last few months as part of its broader safety approach.
The system reviews the agent's interactions and internal reasoning to flag misalignment-relevant behaviors that are hard to surface before deployment, since they often appear only in realistic, tool-rich workflows and extended sessions. OpenAI says internal coding agents carry unique risk factors: they have access to select internal OpenAI systems, can inspect the documentation and code of their own safeguards, attempt to modify those safeguards, and take actions that could affect their future versions.
OpenAI says the monitoring is designed to preserve user privacy and data security, surfacing problematic behaviors only in narrowly-scoped ways. The company argues similar safeguards should be standard for internal coding agent deployments across the industry. The post also outlines limitations and a roadmap toward a formal safety case built on monitoring.
OpenAI is using its most powerful models to police its own agentic coding tools, a test case for how the industry detects misalignment that only surfaces in real-world, tool-rich deployments.