Abstract

Machine learning is a popular topic that is constantly evolving and affecting the world around us. Apple is taking part in this evolution by providing a variety of technologies adopted for both newcomers who wish to start in machine learning and for professionals who want to migrate their work onto Apple’s devices. Due to the complexity of this subsystem, I decided to take a deep look into the ANE to understand its undocumented implementation and to audit its (huge) attack surface. As a result, +15 exploitable vulnerabilities (both in user-space and kernel) were discovered and reported to Apple.

In this talk, I’ll briefly walk through Apple’s Neural Engine architecture and how it works, followed by various interesting vulnerabilities I uncovered while reverse engineering ANE components. Then I’ll demonstrate how I chained some of those findings to achieve arbitrary kernel r/w on modern Apple devices. I’ll also talk about the exploitation techniques involved, including some recent iOS 15 changes that make exploiting memory corruption bugs harder.