LLM Vulnerability Detection with Interprocedural Context Across Multiple Languages

arXiv·medium signal

First empirical study evaluating whether providing caller/callee context to LLMs improves vulnerability detection across function boundaries. Most prior work evaluates LLMs on single-function vulnerabilities, missing data and control flows that span multiple functions. Tests multiple frontier LLMs with and without interprocedural context, measuring both detection accuracy and cost — directly relevant to security-focused AI coding agents.