The vulnerability to watch out for would be the model being trained to not redact some specific information.