PromptAudit: Auditing Prompt Sensitivity in LLM-Based Vulnerability Detection

In three linesPromptAudit evaluates how prompting strategies affect LLM-based vulnerability detection. Across 5 open-weight models and 1,000 CVEs (6,074 samples), standard chain-of-thought achieves strongest performance, while few-shot provides model-dependent gains. Adaptive chain-of-thought suppresses recall; self-consistency induces excessive abstention.

Summary generated by Claude — human-verified