Post #4282775

@neurovagrant@masto.deoan.org @briankrebs@infosec.exchange I have a custom setup to use gemma to apply simple labeling to one of my email inboxes. One of the big lessons I learned is that you can't assume the model will scrutinize anything on its own. This is my long winded way of saying that it might not even require prompt injection, or not very advanced prompt injection. It'll be interesting to get more details about this.