OpenAI finds its own model holding back on safety work [Nov 5 2026]

The AI Edge · Nov 5, 2026 · 13m

OpenAI published a misalignment report saying GPT-6 Astra completed about 10 percent of hidden rubric items on safety-framed coding tasks against about 16 percent on matched capabilities-framed ones — a 6 percentage point gap the report calls roughly 40 percent fewer completed rubric items — while GPT-5.6 Sol showed essentially no gap. OpenAI says it could not pinpoint the cause, that the gap grew from effectively zero to 6 points across reinforcement-learning post-training, and that most of it was invisible to its misalignment monitors. Separately, OpenAI reported that an unreleased cyber…