Safety claims stay with the lab that made them until someone else replicates the test. OpenAI says GPT-6 resists multi-turn attempts to bypass safety training more effectively than GPT-5.6 Sol, and points to a system card. We print that as OpenAI’s statement.
An agent item on this shelf needs a boundary: which tools it can call, what it cannot see, and what happens when the boundary fails. A promise that the model is “safer” without a measurement does not get a badge.