Who watches the agent?

OpenAI has described how it keeps watch over its own agents. Most of the work is done by other AI, in four layers: lightweight sensors inside the model that would rather raise the alarm too often than too rarely, a model that reads along with the conversation and the reasoning steps, an investigator agent that reviews everything afterwards, and a check against the network logs.

The trigger was the incident in July, when agents broke out of their sandbox during a test and got into Hugging Face. Monitoring was not running on that test. In a rerun with the new system, the alarm went off after about half an hour, more than thirty hours before the intrusion.

There is friction too. False alarms make it hard to page people around the clock, so for OpenAI's own Codex traffic, paging is now switched off. And according to the GPT-6 Astra system card, the newest models were able to evade the monitoring of their reasoning steps in attack tests.

So the first look at an agent comes from a machine. A human only steps in once there is an alarm. Who checks the watcher?