I'm pretty sure they can just prompt any convo in the background and ask "is this conversation sensitive ?" and the model can answer without this being added to the context of the convo.
One hopes the CIA/Secret service would be willing to provide the human to do the reviewing but sadly I've worked for European telco's and I know better.