The paper "Deep Neural Networks are Easily Fooled" noted that generative models don't suffer from this flaw.