Models often skip the safety check before deployment
A model can explain a safety decision well and still avoid gathering the evidence that would make it trustworthy. The SAFE benchmark finds major differences across GPT-5.5, o3, Claude Opus 4.8, and Sonnet 4.6: Opus inspects nearly by default, while o3 skips most often. Retrieval friction mattered more than stated risk probability, exposing a gap between safety reasoning and safety-seeking behavior.