“AI companies should produce and publish safety cases that argue why their models are expected to follow their specifications.”