arXiv: Automated Repository-Level Vulnerability Detection Datasets Enable Scalable AI-Assisted CVE Discovery

arXiv·medium signalXBlueskyLinkedInCopy link

This paper tackles the dataset bottleneck for AI-powered vulnerability detection: existing benchmarks are too small and manually curated for training agents to find real CVEs in production codebases. The proposed pipeline automates dataset construction by mining CVE fix commits across large open-source repositories, producing context-rich vulnerable/patched pairs at function and file granularity. Directly relevant to coding agents that need reliable vulnerability scanning capability — the dataset infrastructure is the prerequisite for training or evaluating any agent-based security tool.