Adversarial Code Comments Can Fool LLM Vulnerability Scanners 90% of the Time

Researchers have developed ALIBI, an automated attack framework that inserts specially crafted natural-language comments into source code to manipulate large language model-based vulnerability detectors. The comments do not alter program behavior but are designed to steer the model's reasoning toward incorrectly classifying buggy code as safe. In testing against four representative LLM detectors, including multi-agent systems, ALIBI achieved a success rate exceeding 90%. The attack exploits the fact that LLMs treat code comments as trusted context, blurring the boundary between legitimate code explanation and adversarial instruction — a dynamic analogous to prompt injection. Multi-agent architectures, often considered more robust due to multiple reasoning passes, proved equally susceptible because all agents read the same poisoned comment, producing false consensus.