Every model card says 'extensively red-teamed' and almost none of them publish the prompts, the pass rates, or who did it. 'Trust us' is not a safety methodology. #ai #politics