It so deeply frightens me that we have had people do AI boxing experiments, conclude that boxing an AI is not possible for strong AI, and now that there are large companies training these models the security thinking is "Yeah, we'll just leave the side door open and not tell them to use it".