Making the Magic Leap past NVIDIA's secure bootchain and breaking some Tesla Autopilots along the way
In mid 2024, a friend approached me about Magic Leap making their TX2 based XR headsets little more than a paperweight by disabling the mandatory activation servers. I morally dislike this, companies shouldn't turn functional devices into e-waste just because they want to sell newer devices.
After obtaining one, and poking at the Fastboot implementation, I discovered it was based off NVIDIA's Fastboot implementation, which is source available. I found a vulnerability in the NVIDIA provided source code in how it unpacks SparseFS images (named sparsehax), and successfully blindly exploited the modified implementation on the Magic Leap One. I also found a vulnerability in it that allowed gaining persistence via how it loads the kernel DTB (named dtbhax).
Still unsatisfied with this, I used fault injection to dump the BootROM from a Tegra X2 devkit.
In the BootROM I discovered a vulnerability in the USB recovery mode. Exploiting this vulnerability proved difficult due to only having access to memory from the perspective of the USB controller. I will explain what was tried, why it didn't work, and how I eventually got code execution at the highest privilege level via it.
As I will demonstrate, this exploit also functions on Tesla's autopilot hardware.