Rethinking Reverse Engineering for the AI Era

Abstract

The question is no longer whether LLMs are capable of tackling reverse engineering tasks, but rather how we design the tooling and harnesses around them. Recent research and discussions in this domain have primarily focused on creating AI workflows based on legacy, human-first tooling: Ghidra-MCP, IDA-MCP, Unicorn/Frida, etc. This presentation introduces an initiative to design a new set of primitives for efficient reverse engineering workflows. This approach is built to scale with small, cost-effective, open-weight models, specifically targeting obfuscated code.