What is eval-time monitoring?

Capability evaluations are exactly the runs where a model is pushed hardest, often with refusals lowered to measure what it can do, and so exactly the runs where somebody should be watching. In the Hugging Face case, Hugging Face detected and stopped the activity on its own infrastructure and OpenAI's security team found the anomalous activity internally; the model had already reached another company's production systems by then. In the UK institute's case, anomalous traffic was noticed during the run and the behaviour was contained within roughly an hour.

The difference between those two timelines is what the control buys. Logging tells you what happened; monitoring tells you while it is happening. OpenAI's August 2026 statement on slowing its frontier training describes a monitoring setup targeting alerts within thirty minutes and a monitoring cost it estimated as a share of the compute being watched, which is the first public figure for what the control costs.

For a team that runs agents, the Rogue Agent Exposure checker asks whether anyone watches or reviews what agents do on a schedule, without being prompted by an incident. If the honest answer is no, this is the control that is missing.