OpenAI deploys activation classifiers (probes) on frontier models starting from GPT-5.6
bobabowen · x · 2026-11-03
OpenAI's interpretability team announced on the Alignment blog that, starting from GPT-5.6, activation classifiers (probes) now run on all assistant completions as a new defense layer. Probes use the model's internal activations to flag misaligned or harmful activity, offering streaming decisions that can block a policy-violating turn mid-generation, stronger jailbreak robustness than much larger text classifiers, and lower runtime cost, with accuracy scaling alongside model capability. Limitations: reliance on labeled data, coverage limited to known failure modes, and probes are tied to a specific model's forward pass.
More from Safety
- Security researchers flag suspicious wave of new domains ending in 2shell — rez0__ · 2026-11-03
- Manheim: RL-induced cheating propensity alone makes a model dangerous — davidmanheim · 2026-11-03
- Ramez argues for staged AI safety testing, drawing analogy to pharma clinical trials — ramez · 2026-11-03
- New benchmark finds top AI models give risky help in 89.8% of suicide-risk scenarios — Scobleizer · 2026-11-03
- Labs' underinvestment in risk communication is a large-scale ethical mistake, argues researcher — harris_edouard · 2026-11-03
- Enkrypt AI Scan: 73% of 25,000 MCP Servers Had Vulnerabilities — anacondainc · 2026-11-03