OpenAI details internal safety monitors
OpenAI described activation classifiers that monitor internal model activity and can interrupt responses or block tool calls for further review.
Waiting until an agent finishes acting can be too late to intervene. OpenAI’s activation classifiers are designed to monitor model activity as a response is being generated.
What happened
OpenAI disclosed its approach to activation classifiers, also called probes, on 2 November. These lightweight models examine internal model activity and score completions token by token.
When a score crosses a threshold, the system can pause the response and escalate it to a reasoning model for review. Tool calls remain blocked until they are marked safe. OpenAI has deployed probes as one component of a wider monitoring system alongside other review and response mechanisms.
Why it matters
Monitoring during generation creates an opportunity to intervene before a tool action completes. Using internal activity also offers a different signal from examining the final text alone.
The method still depends on labelled examples of behaviour that the monitoring system is trained to recognise. It cannot be assumed to detect every unfamiliar or poorly defined risk.
The bigger picture
As AI systems gain more ways to act, safeguards need to operate across both model output and tool execution. This disclosure provides a concrete example of how those controls can be connected.
The probes also have operational constraints: they require retraining for new base models, and performance can change as model weights change. Their usefulness therefore depends on continuing evaluation and maintenance as well as the initial training method.