Time Table

As Android relies more on dedicated AI hardware, its software becomes a critical attack surface. During my NUS master’s research, I discovered CVE-2026-57016, a privilege escalation flaw in the Google Pixel EdgeTPU driver. This talk explores a practical question: how well can LLMs analyze an unexplored attack surface?

We will walk through attack surface mapping, finding the bug, and writing the exploit. The main takeaways include learning how to root the Pixel using this vulnerability and examining a completely novel exploitation method built strictly for this bug.

I will also evaluate AI’s performance. While LLMs aided in reverse engineering, they lacked security intuition. I will explain exactly why the LLM initially missed the bug, how I manually directed it to the flaw, and why most models failed to write the full exploit due to safety guardrails and rigid thinking. Finally, I will share the specific strategies I used to overcome these AI limitations.