# OpenAI describes training activation classifiers to detect misaligned model behavior

> Original title: Training activation classifiers for safety and alignment

- Source: OpenAI Alignment Research Blog
- Published: 2026-11-02T20:00:00.000Z
- Bad Influence: https://badinfluence.news/items/k781u0bgeie5weed3tv34tnmi
- Original: https://alignment.openai.com/training-activation-classifiers/

## Summary

OpenAI says it has added activation classifiers, or probes, that run on all assistant completions starting with GPT-5.6. The probes use a model's internal activations to flag misaligned or undesirable activity, can stop a policy-violating turn midway, and are more robust to jailbreaks than external text classifiers. The post also reports that probes must be retrained for new base models and that their accuracy tends to rise with model capability.

## Why it matters

The post explains how OpenAI's activation probes detect misaligned activity, including training choices, transfer limits, and jailbreak robustness, useful for understanding white-box safety monitoring.