Argus: Reorchestrating Static Analysis via a Multi-Agent Ensemble for Full-Chain Security Vulnerability Detection

1. Problem Statement

The paper addresses the limitations of current Static Application Security Testing (SAST) tools and existing Large Language Model (LLM) approaches in detecting software vulnerabilities.

• Limitations of Traditional SAST: Tools like CodeQL and Infer rely on rule-based symbolic analysis and taint tracking. While effective for known patterns, they struggle with:
  • High False Positives/Negatives: Due to incomplete source/sink coverage and inability to handle complex language features (e.g., reflection, multi-threading).
  • Lack of Flexibility: They fail to detect novel or system-specific flaws.
  • Isolated Analysis: They often treat code repositories as "isolated silos," ignoring third-party dependencies and supply chain risks.
• Limitations of Current LLM-Based Approaches: Existing methods typically use LLMs as assistants to traditional tools or for single-pass inference. This leads to:
  • Hallucinations: Generating false vulnerabilities.
  • Shallow Reasoning: Inability to perform deep, multi-step data flow analysis.
  • Poor Integration: Failure to effectively combine LLM reasoning with reachability analysis (e.g., handling disrupted data flows).
  • High Cost: Excessive token usage without proportional gains in detection accuracy.

Core Question: How can we design an LLM-SAST integration that fully leverages the reasoning potential of LLMs while mitigating their drawbacks and integrating effectively with existing tools?

2. Methodology: The Argus Framework

The authors propose Argus (Agentic and Retrieval-Augmented Guarding System), a paradigm shift from "LLM-assisted SAST" to an "LLM-centered SAST" workflow. Argus is a multi-agent framework designed specifically for vulnerability detection.

Key Architectural Components

Argus orchestrates the detection process through three core novelties:

A. Full Supply Chain Analysis (RAG-Enhanced Sink Scanning)
Instead of analyzing code in isolation, Argus parses project dependencies (e.g., pom.xml) and uses Retrieval-Augmented Generation (RAG) to gather vulnerability intelligence.

• Data Sources: It queries authoritative databases (NVD, OSV, GHSA, Snyk) and community sources (GitHub issues, forks).
• Structured Output: It generates structured vulnerability reports including descriptions, severity, affected versions, and CVE IDs.
• PoC Generation: A dedicated PoC Agent uses the ReAct (Reasoning + Acting) paradigm to iteratively verify exploitability, generate Proof-of-Concept code, and create patches. This validates sinks before deep analysis begins.

B. Multi-Agent Collaboration
The SAST pipeline is decoupled into specialized agents:

• Dependency Scanning Agent: Identifies active dependencies.
• CVE Retrieval Agent: Gathers external vulnerability data.
• PoC Generation Agent: Verifies exploitability.
• Sink Scanning Agent: Identifies dangerous functions (sinks).
• Flow Review Agent: Audits data flow correctness.

C. Re3 Methodology (Retrieval, Recursion, Review)
To solve the "stealthy interruption" problem where data flows are broken by complex code features, Argus employs a novel Re3 workflow for data flow analysis:

1. Retrieval: Uses CodeQL to perform forward data flow analysis from sources to identified sinks.
2. Recursion: If a direct path is blocked (e.g., by reflection), the system performs backward analysis from the sink to find upstream function calls. These upstream nodes are treated as new "surrogate sinks," and forward analysis is re-attempted to bridge the gap.
3. Review: A dedicated LLM agent performs a three-step audit:
   • End-to-End Reachability: Checks for control flow interruptions (if/else, try/catch).
   • Hop-by-Hop Analysis: Verifies sanitization, encoding, or type casting at every step.
   • Document Export: Generates a structured report for human review.

3. Key Contributions

1. First LLM-Centered SAST Framework: Argus is the first framework to treat the LLM agent as the primary orchestrator, with SAST tools and external data serving as supplemental inputs.
2. Supply Chain Integration: It is the first to integrate full supply chain analysis (dependencies) directly into the static analysis workflow using RAG.
3. Novel Re3 Strategy: Introduces a recursive, retrieval-based approach to data flow analysis that overcomes the limitations of single-pass symbolic analysis.
4. Zero-Day Discovery: The framework successfully identified critical zero-day vulnerabilities in real-world codebases, which were assigned CVE numbers.

4. Experimental Results

The authors evaluated Argus on seven popular, industrial-level Java open-source repositories (e.g., PublicCMS, JeecgBoot, KeyCloak) containing millions of lines of code.

• Baselines: Compared against CodeQL (standard SAST tool) and IRIS (a recent LLM-based Java vulnerability detector).
• Performance Metrics:
  • Vulnerability Detection: Argus significantly outperformed baselines. While CodeQL and IRIS often detected 0 vulnerabilities in these specific test cases (due to lack of context or rule coverage), Argus detected 3 to 6 verified vulnerabilities per repository.
  • Sink Discovery: Argus identified significantly more "stealthy" sinks than CodeQL, largely due to the RAG-enhanced discovery of dependency-related sinks.
  • Token Efficiency: Despite using LLMs, Argus demonstrated lower token consumption per detected vulnerability compared to naive LLM approaches, proving its efficiency.
  • Zero-Day Impact: The system discovered potential zero-day vulnerabilities that were previously unknown, leading to official CVE assignments.

Key Finding: The ablation study showed that sinks discovered via the RAG/Community approach yielded significantly more vulnerabilities than those found by static tools alone, validating the importance of supply chain context.

5. Significance

• Paradigm Shift: Argus moves the field from "LLMs as a helper" to "LLMs as the central reasoning engine," fundamentally changing how SAST is orchestrated.
• Industrial Viability: By reducing false positives and operational costs (token usage) while increasing true positive detection, Argus offers a practical solution for industrial deployment.
• Security Impact: The ability to automatically mine zero-day vulnerabilities and generate PoCs provides a proactive defense mechanism, potentially preventing catastrophic breaches before they are exploited.
• Open Source: The authors plan to release the framework to benefit both academia and industry, fostering further research in agentic security testing.

In conclusion, Argus represents a significant advancement in automated security testing by effectively combining the contextual reasoning of LLMs, the precision of symbolic tools (CodeQL), and the breadth of supply chain intelligence through a multi-agent, retrieval-augmented architecture.