A dangerous question would have to be along the lines of "Could I use unobtanium with the Tony Stark process to produce explosives much more powerful than nuclear weapons?" so that the question itself contains some insight that gets you closer to doing something dangerous.
Perhaps the reason for not publishing the questions is twofold: 1) they want a universal jailbreak that can get the model to answer any "bad" question. 2) they don't want bad publicity when someone not under NDA jailbreaks their model and answers their question