train them to accomplish the tasks necessary to implement each attack strategy.

It seems to me misguided to take these potentially dangerous models then train them to display dangerous behaviours. I prefer using proxy measures.