Abstract

LLMs are changing the pace of vulnerability research. At Pwn2Own Berlin, the volume of registrations was high enough that ZDI had to adjust its registration rules, reflecting how quickly researchers can now find and develop bugs with AI assistance. But finding a vulnerability is only the first step. Understanding how far it can be pushed is a different challenge.

Our research began with a simple WebKit vulnerability visible in plain sight in the source code. Unfortunately, it turned out to be a duplicate of an earlier report, and its impact had been underestimated as merely a cross-origin violation. We later found a related vulnerability nearby and showed that this assessment had stopped far too early. To complete the exploit chain, we also needed a separate PAC-bypass vulnerability. Finding it required careful reasoning about multiple primitives and a willingness to keep exploring when LLMs failed to produce a useful answer. By combining these pieces, we built a complete Safari exploit chain for Pwn2Own.

This talk is about what happens after the obvious finding. LLMs accelerate source review, suggest hypotheses, and quickly surface possible impacts, but they are less reliable when exploitation requires deeper investigation, architectural understanding, and the assembly of seemingly unrelated primitives. These limitations may not last forever. But regardless of how mighty LLMs become, the willingness to challenge the answer and take the analysis one step further remains the heart of hacking.