While I understand the premise I think this is a highly flawed way to operate these tools. I wouldn't want to have someone with my personal data (whichever part) that might give it to anyone who just asks nicely because the context window has reached a tipoff point for the models intelligence. The major issue is a prompt attack may have taken place and you will likely never find out.