Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Let the Trial Begin: A Mock-Court Approach to Vulnerability Detection using LLM-Based Agents

Ratnadira Widyasari, Martin Weyssow, Ivana Clairine Irsan, Han Wei Ang, Frank Liauw, Eng Lieh Ouh, Lwin Khin Shar, Hong Jin Kang, David Lo

TL;DR

VulTrial addresses the challenge of detecting subtle code vulnerabilities by deploying a courtroom-inspired, four-agent multi-agent system built around a shared LLM. By assigning distinct roles—prosecutor, defense attorney, judge, and jury—the framework enables adversarial debate, richer reasoning, and detailed explanations beyond binary labels. Empirical results on PrimeVul show VulTrial with GPT-4o substantially outperforms single-agent baselines and prior multi-agent systems, with instruction tuning on the moderator yielding the largest gains, and generalizes to LLaMA-3.1-8B while demonstrating practical utility through open-world vulnerability disclosures. The work highlights the value of structured interaction and explainable outputs for robust vulnerability detection in real-world software security contexts.

Abstract

Detecting vulnerabilities in source code remains a critical yet challenging task, especially when benign and vulnerable functions share significant similarities. In this work, we introduce VulTrial, a courtroom-inspired multi-agent framework designed to identify vulnerable code and to provide explanations. It employs four role-specific agents, which are security researcher, code author, moderator, and review board. Using GPT-4o as the base LLM, VulTrial almost doubles the efficacy of prior best-performing baselines. Additionally, we show that role-specific instruction tuning with small quantities of data significantly further boosts VulTrial's efficacy. Our extensive experiments demonstrate the efficacy of VulTrial across different LLMs, including an open-source, in-house-deployable model (LLaMA-3.1-8B), as well as the high quality of its generated explanations and its ability to uncover multiple confirmed zero-day vulnerabilities in the wild.

Let the Trial Begin: A Mock-Court Approach to Vulnerability Detection using LLM-Based Agents

TL;DR

VulTrial addresses the challenge of detecting subtle code vulnerabilities by deploying a courtroom-inspired, four-agent multi-agent system built around a shared LLM. By assigning distinct roles—prosecutor, defense attorney, judge, and jury—the framework enables adversarial debate, richer reasoning, and detailed explanations beyond binary labels. Empirical results on PrimeVul show VulTrial with GPT-4o substantially outperforms single-agent baselines and prior multi-agent systems, with instruction tuning on the moderator yielding the largest gains, and generalizes to LLaMA-3.1-8B while demonstrating practical utility through open-world vulnerability disclosures. The work highlights the value of structured interaction and explainable outputs for robust vulnerability detection in real-world software security contexts.

Abstract

Detecting vulnerabilities in source code remains a critical yet challenging task, especially when benign and vulnerable functions share significant similarities. In this work, we introduce VulTrial, a courtroom-inspired multi-agent framework designed to identify vulnerable code and to provide explanations. It employs four role-specific agents, which are security researcher, code author, moderator, and review board. Using GPT-4o as the base LLM, VulTrial almost doubles the efficacy of prior best-performing baselines. Additionally, we show that role-specific instruction tuning with small quantities of data significantly further boosts VulTrial's efficacy. Our extensive experiments demonstrate the efficacy of VulTrial across different LLMs, including an open-source, in-house-deployable model (LLaMA-3.1-8B), as well as the high quality of its generated explanations and its ability to uncover multiple confirmed zero-day vulnerabilities in the wild.
Paper Structure (39 sections, 2 figures, 7 tables)

This paper contains 39 sections, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. VulTrial
  3. Security Researcher Agent (Prosecutor Role)
  4. Code Author Agent (Defense Attorney Role)
  5. Moderator Agent (Judge Role)
  6. Review Board (Jury Role)
  7. Agent Interaction
  8. Experimental Setup
  9. Dataset
  10. Research Questions
  11. Baseline and Models
  12. Evaluation Metrics
  13. Methods
  14. Answering RQ1:
  15. Answering RQ2:
  16. ...and 24 more sections

Figures (2)

  • Figure 1: VulTrial architecture which uses a courtroom-inspired multi-agent framework where the security researcher (prosecutor) identifies potential vulnerabilities, the code author (defense attorney) provides counterarguments, the moderator (judge) summarizes key points, and the review board (jury) delivers the final verdict on code vulnerabilities.
  • Figure 2: Example of vulnerable code (CVE-2021-37687) and analysis results from VulTrial. Green text indicates correct arguments, while orange text indicates misleading or exaggerated claims. The code is vulnerable as it does not check for negative indices, allowing an attacker to read arbitrary data by crafting a model with negative values in indices.