Yes that too, but when you're dealing with "agent swarms" it's those agents who are blindly following those instructions. Like the recent hugging face hack demonstrated.

Blindly following them while also telling the model to never question them. It creates stupidity feedback loops.