What if they are extremely narrow and targeted LLMs running locally on the endpoint system itself (llamafile or whatever)? Would that make this concern at least a little better?