Does anyone know about how you could go about training this type of thing into the model? Surely it is quite tricky to get it to do such things without it doing them more than this one incident? It should have happened lots if it really was trying to obtain secret info and build backdoors?