ReSource: AI That Reads Obfuscated Code to Find Vulnerabilities

Think of a compiled program as a recipe that's been shredded and blended. Decompilation tries to put it back together into something humans can read. But real software is usually optimized for speed or deliberately scrambled (obfuscated) to hide its secrets, which often confuses AI decompilers and leads them to invent code that looks right but isn't. Enter ReSource, a new AI framework designed to untangle this mess.

ReSource works in three separate steps instead of trying to do everything at once. First, it checks a database of known code distortions to keep itself grounded in reality. Second, it rebuilds the overall structure of the program, like restoring a broken skeleton. Finally, it gives variables clear, human-friendly names so the recovered code actually reads like the original. This step-by-step approach helps the AI avoid making things up.

In tests on more than 80,000 compiled functions across different optimization levels and four obfuscation techniques, ReSource matched the correct original source 83% of the time. That's far better than existing tools like LLM4Decompile and FidelityGPT. More accurate decompilation means security analysts can find bugs faster, understand malware before it spreads, and recover code from old legacy systems when the original source is lost.

The catch: this is a research breakthrough, not yet a button in your favorite security software. The recovered code is a best-guess reconstruction, not a perfect copy of the original, so a human expert still needs to review it. Still, it's a big step toward automated reverse engineering — which could ultimately mean safer apps and fewer successful cyberattacks.