Have adversarial attacks to Neural Networks been a real threat so far?

definitely yes as real threat. whether or not successfuly exploited - some researchers have shown potential exploitations, I can't vouch for any "in the wild" exploits but I am certain this is something under active research.

I myself planned a proof of concept for such an attack, although it has proven trickier than expected. I'm willing to collaborate with someone for the sake of publishing a nice article on the matter (don't expect much, it's not a "real" threat but a small PoC that might get some buzz and be used as a nice case study for future research)