Comment by Sambhav Maheshwari

AI developers should aim to build a “safety case,” i.e., a structured argument, supported by evidence, that the AI system is acceptably safe. A safety case need not be a formal proof of safety; rather, it is a comprehensive assessment of the available evidence, organized to make the reasoning transparent and contestable. […] For risks from internal models, this requires separate arguments for the autonomous AI misbehavior and the rogue insider threat pathways.