FuzzingBrain V2: Multi-Agent LLM System Achieves 90% Vulnerability Detection Rate and Discovers 29 Zero-Day Flaws

Summary

FuzzingBrain V2 is a multi-agent LLM system for automated vulnerability discovery and reproduction in software. It addresses three key challenges: high false positive rates in LLM-generated reports, suboptimal granularity for vulnerability localization, and difficulty reasoning about complex cross-function dependencies. Built on Google's OSS-Fuzz, it introduces Suspicious Point (a control-flow-based abstraction), logic-driven hierarchical function analysis with dual-layer fuzzing, and MCP-based static/dynamic analysis tools. In the AIxCC 2025 Final Competition, it achieved a 90% detection rate (36/40 vulnerabilities), and in real-world deployment discovered 29 zero-day vulnerabilities across 12 open-source projects, with 2 assigned CVE IDs.