A new preprint on arXiv examines how AI agents, which have shown promise in automating vulnerability discovery, can also be turned toward attacking software systems at scale. The authors argue that today's software needs defensive mechanisms to level the playing field against such automated threats.
The paper, titled "Natural Code Transformation Confuses Agents in Vulnerability Analysis," appears to investigate whether applying natural code transformations—changes that preserve functionality while altering the code's surface—can disrupt an AI agent's ability to analyze vulnerabilities. The abstract indicates that these transformations do confuse agents, though the full details of the experiments are not available in the provided excerpt.
The authors frame this as a potential defensive strategy: by making code harder for AI agents to analyze, systems could slow down automated attacks. The study is preliminary, and the abstract is truncated, so the specific methods and results remain unclear from the source alone.