Enhancing Software Vulnerability Analysis via Large Language Models: A Systematic Review

In the current era of AI advancement, large language models (LLMs) have achieved remarkable breakthroughs in natural language understanding and code generation, and have been rapidly adopted in the field of software security analysis. This paper presents a systematic review of the research advancements in applying LLMs to vulnerability analysis tasks such as binary code analysis, static and dynamic vulnerability detection, and vulnerability validation. Studies show that the strong contextual understanding capabilities of LLMs enable the identification of common vulnerability types across multiple programming languages. Moreover, LLMs can act as intelligent attack agents to perform end-to-end penetration testing. The integration of LLMs with traditional static analysis techniques—such as symbolic execution and taint analysis—also demonstrates significant potential for improving the accuracy of security analysis. However, the application of LLMs in software security still faces several challenges, including high false positive rates, severe hallucination issues, limited context modeling capacity, and substantial resource consumption. This survey aims to provide researchers with a comprehensive technical overview, summarizing the latest achievements in leveraging LLMs for software security analysis, discussing their strengths and limitations, and outlining future research directions.