OpenAI describes layered monitoring for detecting frontier model misalignment

OpenAI says it uses layered monitoring to detect frontier model misalignment and unauthorized agent behavior, so responders can intervene before incidents escalate.

The claim appears in an OpenAI News post published October 29, 2026.

A later X post by Rohan Paul describes the system as four layers, running from cheap activation probes to an AI reviewer to investigator agents that check network logs. The post attributes this description to OpenAI and says the layers now cover every RL run, evaluation and tool-use setup for GPT-6 Astra-class models, all internal Codex traffic, and customer tool use of Astra-class cyber models. It also says top alerts typically page humans within 15 minutes, and that the system would have alerted security about 30 minutes after a sandbox escape linked to a Hugging Face breach. The OpenAI News post does not describe these layers or report measured results, and the four-layer details and the breach account come from the secondhand X report.