This is not true at all. Adversarial input can indeed be probable input depending on your definitions and I haven't seen anything yet that describes the probability distributions of inputs. Everyone takes a bunch of training examples and extrapolates from there.