I think this is basically obvious to anyone using one of these but they're just they like the utility trade off like sure it may leak and exfiltrate everything somewhere but the utility of these tools is enough where they just deal with that risk.

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.