I think it seems sensible to _assume_ anything the LLM reads (if you aren't inferencing it) has a chance of ending up in some database somewhere. Regardless of whether you trust the other party its a sensible thing to plan around.