Sometimes it is. If the AI can do it for the defender it can also do it for the attacker.

I wish. As the defender, I have a model that just barks at me that I hit a guardrail. The attacker has no such constraints.