Seems like they could have potentially gotten access to the weights if they weren’t concerned about not doing crimes.

Like how forums are always hosted on different servers from monorepos, so therefore it's impossible to hack the OpenAI monorepo from an OpenAI forum?

Maybe not the current models, maybe not this year. But even a almost perfectly aligned model will misbehave one day.