Assessing LLMs in malicious code deobfuscation of real-world malware campaigns

Identificador:  imarina:9379058

Handle: https://hdl.handle.net/20.500.11797/imarina9379058

Autors:  Patsakis, Constantinos; Casino, Fran; Lykousas, Nikolaos

Resum:
The integration of large language models (LLMs) into various cybersecurity pipelines has become increasingly prevalent, enabling the automation of numerous manual tasks and often surpassing human performance. Recognising this potential, cybersecurity researchers and practitioners are actively investigating the application of LLMs to process vast volumes of heterogeneous data for anomaly detection, potential bypass identification, attack mitigation, and fraud prevention. Moreover, LLMs' advanced capabilities in generating functional code, interpreting code context, and code summarisation present significant opportunities for reverse engineering and malware deobfuscation. In this work, we comprehensively examine the deobfuscation capabilities of state-of-the-art LLMs. Specifically, we conducted a detailed evaluation of four prominent LLMs using real-world malicious scripts from the notorious Emotet malware campaign. Our findings reveal that while current LLMs are not yet perfectly accurate, they demonstrate substantial potential in efficiently deobfuscating payloads. This study highlights the importance of fine-tuning LLMs for specialised tasks, suggesting that such optimisation could pave the way for future AI-powered threat intelligence pipelines to combat obfuscated malware. Our contributions include a thorough analysis of LLM performance in malware deobfuscation, identifying strengths and limitations, and discussing the potential for integrating LLMs into cybersecurity frameworks for enhanced threat detection and mitigation. Our experiments illustrate that LLMs can automatically and accurately extract the necessary indicators of compromise from a real-world campaign with an accuracy of 69.56% and 88.78% for the URLs and the corresponding domains of the droppers, respectively.