AI-Assisted Vulnerability Research: Building a Binary Audit Agent

AI-Assisted Vulnerability Research: Building a Binary Audit Agent - kdz

Vulnerability research on closed-source binaries is slow, specialist work. It means pulling an executable into a disassembler, recovering its structure by hand, tracing attacker-controlled data flows, and proving that suspicious code fragments are reachable and exploitable. What if an AI agent could help drive that entire pipeline?

This talk walks through my attempt to design such an agent using a custom Claude Code plugin (very much a work in progress and a learning opportunity for me). The agent orchestrates a workflow that follows a target end to end: QA'ing the decompilation, sweeping for dangerous code, reasoning over the code to build a whole-binary flow map that traces attacker-controlled input through to dangerous operations, and then attempting to craft and verify a payload that actually triggers the bug.

Location
Room 080.02.002 at RMIT (Building 80) - 445 Swanston St, Melbourne

Discord
Discord Invite
Discord Event Info