Recoleta
GitHub
Source note

FuzzingBrain V2: A Multi-Agent LLM System for Automated Vulnerability Discovery and Reproduction

Authors
Ze Sheng, Zhicheng Chen, Qingxiao Xu, Kewen Zhu, Jeff Huang
Source
arXiv
Published
Collection
Software Intelligence
LLM AgentsVulnerability DiscoveryFuzzingCode IntelligenceSoftware Security
Open arXivMarkdown source

Summary

FuzzingBrain V2 is a multi-agent LLM system that finds C/C++ vulnerabilities and verifies them with OSS-Fuzz crashes. Its main claim is that LLM analysis becomes more useful when every reported issue must produce a reproducible fuzzer input.

Problem

  • LLM vulnerability reports often have false positives because they lack executable proof-of-concept inputs.
  • Function-level review can hide bugs in large context, while line-level review often lacks enough context to judge exploitability.
  • Bugs with cross-function data flow, indirect calls, state, checksums, or protocol steps are hard for plain static tools and LLM-only review.

Approach

  • The system runs workers for each fuzzer and sanitizer pair, such as libFuzzer with AddressSanitizer, and uses OSS-Fuzz as the verification backend.
  • It introduces Suspicious Points: structured bug candidates tied to control-flow locations, vulnerability type, score, reachability, verification status, and PoC guidance.
  • Specialized agents generate directions, screen functions for Suspicious Points, deduplicate candidates, verify reachability and safety checks, generate PoC inputs, and write reports.
  • It uses a hierarchical search over feature-oriented code regions, then analyzes one function at a time with caller, callee, static-analysis, and dynamic-analysis context.
  • It combines LLM-generated inputs with two fuzzing layers: a global fuzzer for broad exploration and an SP fuzzer that mutates failed PoC attempts while the LLM tries new inputs.

Results

  • On the AIxCC 2025 Final Competition C/C++ dataset, it reports a 90% detection rate: 36 of 40 vulnerabilities found.
  • The abstract reports 29 zero-day vulnerabilities across 12 open-source projects, all confirmed and fixed by maintainers, with 2 CVE IDs assigned.
  • The introduction also claims 41 zero-day vulnerabilities across 19 open-source projects; the excerpt does not resolve this conflict with the abstract.
  • The paper claims 100% reproducibility for confirmed vulnerabilities because a confirmed report requires a sanitizer-detected crash input.
  • The system is designed to run on more than 1,000 OSS-Fuzz integrated open-source projects.

Link