VVF-AI: A Vulnerability Verification Framework Based on AI-Agent

With the continuous deepening of research on big language models in LLM4Cybersecurity, LLM has shown great potential in applications such as vulnerability detection, code repair, and threat intelligence. This study innovatively proposes an AI agent based PoC verification framework (VVF-AI) to address the technical challenge of high false positive rates in vulnerability detection. We manually validate and evaluate four types of vulnerabilities based on the benchmark we constructed. The experimental results demonstrate that LLM can automatically filter false positives and achieve an accuracy rate of 93.1 % in information leakage vulnerabilities. The proposed framework is applicable to ASAT tasks.