Reproduce “Metamorphic Malware Evolution: The Potential and Peril of Large Language Models” on OpenResearch

About this paper

Pooria Madani

Pooria Madani's research at Ontario Tech University explores how Large Language Models (LLMs) can generate diverse, functionally equivalent code variants, revealing their potential for creating highly sophisticated metamorphic malware. The study quantifies this capability, showing leading LLMs like ChatGPT 3.5 Turbo can produce unique and correct solutions over 50% of the time, which impacts future cybersecurity defense strategies.

Key insights

• Current general-purpose LLMs possess an inherent capability for code metamorphism, able to generate multiple syntactically distinct yet semantically identical code variants without explicit training for mutation.
• LLMs can generate solutions using fundamentally different algorithmic approaches (e.g., iterative vs. recursive), surpassing the diversity limitations of traditional rule-based mutation engines.
• The introduced framework and `variation@k` metric provide a systematic methodology for assessing LLMs' potential for generating metamorphic code, enabling proactive development of anti-malware defenses.

Summary by alphaXiv.