Tuesday, November 3, 2026 · ChatGPT & OpenAI
OpenAI · Safety · happened Nov 2, 2026
OpenAI says activation probes have checked every assistant response since GPT-5.6 and can stop output mid-stream
In an Alignment blog post on Nov. 2, OpenAI described lightweight classifiers that read a model's internal activations and score output token by token, so a response or tool call can be halted before it finishes; flagged messages go to a reasoning model for review. OpenAI says they run on internal and external traffic and held up better than text classifiers against the jailbreaks it tested. It notes they only catch behaviors they were trained on and must be retrained for each new base model.
Why it matters
It is a rare detailed look at a monitoring layer running on production ChatGPT traffic.