In its March 2026 description of internal coding-agent monitoring, OpenAI said its monitor reviewed completed interactions and that blocking selected actions before execution was a planned extension.

OpenAI reported a monitoring system that analyzes coding-agent traces and alerts human responders, with its latest form reviewing interactions within 30 minutes of completion. It separately proposed a more synchronous stack that could evaluate and potentially block the highest risk actions before execution. Thus the report supports two distinct observations: a capable model is already used to monitor agents, and the lab is exploring a move from detection after the fact to prevention at a gate. It does not report a measured catastrophic-risk reduction from that proposed gate.