Completion monitors need retraining

An OpenAI alignment research post dated November 2, 2026 says that, starting from GPT-5.6, the most capable models added activation classifiers on all assistant completions. The classifiers score a completion as it is generated. They can interrupt a policy-violating turn midway, block tool calls, and pause the completion while a reasoning model runs a review.

The post says the classifiers are deployed on both internal and external traffic of the most capable models. They are one part of an internal misalignment monitoring system, alongside model-based review, automated investigation, and human response. They complement other safeguards and do not replace them.

The reported experiments say the classifiers show greater robustness to the jailbreaks that were tested than the text classifiers, generally improve as the monitored model becomes more capable, and transfer across many post-training checkpoints. The figures come from internal experiments and are not necessarily representative of the deployed classifiers.

The limits are part of the deployment note. The classifiers use supervised labels and can detect only kinds of activity that can be explicitly trained. They depend on the monitored model's activations, so each new pre-trained base model needs a retrained classifier. Within one base model, performance can degrade as the weights move away from the checkpoint used for training, which is why retraining may be repeated. The later sections describe architecture and training. This column stays with the deployment and the stated limits.