The basic AI security control is “Permission-aware retrieval” and not “broader indexing.” An assistant should only surface what the person asking is already entitled to see.