Exploring the Impact of LLM Assisted Malware Variants on Anti-Virus Detection

In this study, we develop 400 malware variants of four different types of test malware samples by rewriting their original source code with OpenAI's gpt-3.5-turbo model [1], and investigate if these samples can be detected by anti-virus. From this investigation, we find that about half of the malware variant function correctly and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$37.6\pm 33.6$</tex> % of them evade anti-virus detection for each malware type and anti-virus.