Adaptive Agentic Attacks on LLM Vulnerability Detectors via Adversarial Comments (arxiv.org)

A recent study has introduced ALIBI, an automated black-box attack framework designed to exploit vulnerabilities in Large Language Model (LLM)-based vulnerability detectors. As LLMs are increasingly used for critical tasks like vulnerability detection and code review, researchers discovered that adversarial code comments can significantly impact a detector's performance without altering the actual behavior of the code. The framework simulates a coding agent that deliberately injects vulnerabilities and strategically places comments that mislead the detection process. Alarmingly, the study found that existing detectors are highly susceptible to these attacks, with success rates exceeding 90% for real-world vulnerabilities, and reaching 100% in some cases. The implications for the AI/ML community are profound, highlighting a previously overlooked attack vector that threatens security in LLM applications. The iterative nature of ALIBI allows for continuous refinement based on feedback from detectors, making these attacks even more effective. Additionally, while some prompt-level defenses provide minimal resistance, enhancements like architectural isolation and comment sanitization show promise in improving the robustness of vulnerability detectors. This research emphasizes the need for security-aware designs in AI systems that carefully manage the relationship between natural language context and programming logic, underscoring the urgency for advancements in securing AI technologies against adaptive threats.