Is this true? I would have thought all you would need is to give it an input that maps to a 3d surface that's adversarial. There's an extra step in the pre-prep pipeline, but the basic technique is the same - gradient descent on inputs until you derive those that are sufficiently adversarial.

All neural nets are vulnerable to adversarial examples. It's a fundamental property they hold, because they're essentially stacked linear models. So (for example) they get more confident about their predictions when given a sufficiently out-of-domain input - adversarial training is essentially just finding paths that trigger an out-of-domain response.

I don't see how an additional transformation before input precludes that.