I'm far from an expert, but my understanding was that those kinds of adversarial attacks would be prohibitively difficult without full access to the learning system used by that product. If so, there's a vector if you can hack the company and train your own model to generate the adversarial examples to fool it, but it's not a trivial thing