Is sandboxing sufficient to contain rogue agents?
Extrait original
So on this point I’m going to side with the infosec folks. The labs have not been doing containment correctly, and so we can’t really tell if the problem is models or just bad infrastructure.