It's a whole field of research, and it's pretty trivial to generate them for most classes of ML models. It's actually quite difficult to create robust models that DON'T have this problem...