That why you tell them not to harm anyone in their instructions.

At that point AI would be incredibly advanced and adept, I doubt task prioritization will be a issue.

Hit them with the
if(goingToDoBadThing()){
dont();
}