Reproduce “Multi-Agent Taint Specification Extraction for Vulnerability Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jonah Ghebremichael, Saastha Vasan, Saad Ullah, Greg Tystahl, David Adei, Christopher Kruegel, Giovanni Vigna, William Enck and 1 more

SemTaint, a multi-agent system, enhances static analysis for JavaScript by strategically leveraging large language models to resolve dynamic behaviors and model complex dependencies. This approach achieved a 65.43% recall rate on previously undetected npm vulnerabilities and led to the discovery of four new exploitable flaws.

Key insights

  • LLMs can effectively augment traditional static analysis by semantically reasoning about dynamic behaviors and package-specific patterns where static tools fall short.
  • Focusing LLM inference on security-relevant paths and applying demand-driven analysis dramatically reduces the computational overhead, making LLM integration scalable for large codebases.
  • Structuring extracted taint specifications as reusable external predicates allows for modular integration with existing SAST tools and iterative improvement of the underlying data flow graph.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.