As more of our society moves to and relies on computer systems, the software that runs on these systems has become so large and complex that traditional approaches to software security are unable to keep up with it. On the other hand, machine learning prospers on large, diverse data sets. I will discuss our endeavor to apply machine learning to software security. I will show how to transform fuzz testing from a random, ad hoc approach to a principled approach based on machine learning techniques, and demonstrate that this principled approach solves important problems that challenge traditional software testing. I will also discuss how machine learning can benefit from software engineering. I will show how to use fuzz testing to assist code representation learning and to improve program understanding tasks. I will discuss the challenges, opportunities, and open problems in AI-driven software security.