OpenAI describes activation probes that monitor its frontier models token by token, deployed since GPT-5.6
Confirmed
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 4 November 2026
Your AI and this story
- GPT-6 Astra186 days after its cutoff
- Claude Opus 5.5125 days after its cutoff
- Gemini 4 Argon216 days after its cutoff
- Grok 4.7155 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 125 days before it.
Key facts
- Deployed on OpenAI’s most capable models starting with GPT-5.6
- Probes read residual-stream activations, usually from a middle layer; attention-probe architecture
- Trained on up to ~1M messages labelled by reasoning models acting as judges; random initialisation explained ~85% of variance in probe quality
- Accuracy tends to rise with the capability of the monitored model; transfer is good across post-training checkpoints, weaker across pre-training ones
- Limits: detect only behaviours they were trained on; need retraining for each new base model
What happened
Authors Alex Makelov, Xiaojun Xu, Sundaram Ananthanarayanan, Ethan Mcdonald and Tom Dupre la Tour describe the probes in production use. “Probes complement other safeguards rather than replace them,” the post says.
Why it matters
It is one of the first public accounts of interpretability-style monitors running in a major lab’s production stack rather than in research.
Sources
1 source from 1 site. Numbers match the chips in the text.
1 source: 1 primary
Primary
- OpenAI Alignment: Training activation classifiers for safety and alignmentalignment.openai.com, official
Changes
- Filed