New AI Unscrambles Messy Code to Help Catch Hidden Malware

Almost every website you visit runs JavaScript — the instructions that make buttons click, forms submit, and pages load. Companies often deliberately scramble that code (called "obfuscation") to protect their work or make it load faster. Criminals use the exact same trick to hide malicious code inside websites and ads, which is why security analysts spend hours manually untangling it.

This paper, from researchers Rong Feng and Suman Saha, tries to automate that untangling. Most AI attempts treat the job as one big translation: paste the scrambled code in, get clean code out. That usually fails, because real-world deobfuscation happens in steps — peel one layer, look at the result, peel the next. Their approach trains the AI on those intermediate steps (using an existing deobfuscation tool as a guide), but at the moment of use, it produces the cleaned-up program in a single pass.

The results are a real improvement, but modest. On a standard test set, the model produced syntactically valid code — code a computer could actually run — 50% of the time. Direct fine-tuning managed 15%, and an AI given no training managed 25%. Among the outputs that were valid, 80% behaved the same as the original code, meaning the meaning survived the cleanup. Half the attempts still failed, and this is labeled a work-in-progress paper, tested on one benchmark.

For you, the practical upside is eventually faster detection of malicious scripts on websites you visit, and quicker security reviews of the apps you use. The catch: the same technology could help attackers understand and reuse other people's code. Right now, think of this as promising lab work — not something protecting your browser today.