OpenAI deploys activation classifiers in GPT-5.6

OpenAI describes its deployment and methodology for training activation classifiers, or probes, as an additional layer of defense to monitor frontier model safety and alignment. These lightweight ML models use an LLM's internal residual stream activations to classify undesirable activity in a streaming fashion, offering advantages in efficiency and robustness over external text classifiers. The architecture incorporates Attention probes with QV-attention, exponential moving average, and a rolling softmax objective to improve optimization and streaming performance. Open article →