"Adversarial examples generalize across models trained to perform the same task, even if those models have different architectures and were trained on a different training set."