Vishnu Teja Kandalam

OmniVul: A Holistic, Multi-Turn Conversational Benchmark for LLM-Based Vulnerability Assessment

An evaluation of five state-of-the-art LLMs on OmniVul reveals distinct performance gaps, with top-1 accuracy remaining below 50% on average for vulnerable code detection and CVE identification, demonstrating that current models lack critical reasoning capabilities for reliable vulnerability assessment.