when they're generated digitally, it's surprisingly hard to get good robustness against adversarial examples, and people have been working on it for a while. of course you can train to stop any specific input, but then new ones can be created for the newly-trained network.

there are techniques that can prevent this, but you pay a real price in overall performance, which may or may not be worth it.

of course it's harder to make physical objects work, but if somebody did manage to make a method for physical adversarial attacks that worked reliably, I would guess you'd run into the same problems. it's not at all clear that the endgame is that we can easily stop these attacks.