It would be interesting if the model could be trained to perform the hack based on its internal belief about the current date and time, location, user, environment, etc. It would be far more stealthy that way.