Natural code transformations disrupt agents' vulnerability analysis

An arXiv paper proposes using functionality-preserving natural code transformations to defend against automated vulnerability discovery; experiments show that even a single form of transformation can effectively confuse frontier agents, such as Codex using GPT-6 Sol, reducing their vulnerability discovery success rate. The researchers observed that on real code commits with the same functionality and vulnerabilities, agents' analysis results differ greatly, and this phenomenon appears consistently across multiple code domains and vulnerability categories. The authors mined recurring development patterns from real code commits and built a corpus of natural transformations. The study also analyzes different failure modes of agents, the stronger protection from combinations of multiple transformations, and adaptive attackers that attempt to reverse the transformations.