We are still learning how to cope with adversarial machine learning. Security researchers are used to perusing code for vulnerabilities. Now they must learn to find security holes in AI systems composed of millions of numerical parameters.