Multi-Agent LLM System for Automated Vulnerability Discovery and Reproduction
2 hours ago
- #Multi-Agent LLM
- #FuzzingBrain V2
- #Vulnerability Discovery
- FuzzingBrain V2 is a multi-agent LLM system designed for automated vulnerability discovery and reproduction.
- It addresses high false positive rates and lack of reproducible verification in LLM-generated vulnerability reports.
- It introduces 'Suspicious Point', a control-flow-based abstraction for precise vulnerability localization at an optimal granularity.
- The system uses logic-driven hierarchical function analysis and dual-layer fuzzing for enhanced function coverage.
- It incorporates MCP-based static and dynamic analysis tools with context engineering for complex vulnerability reasoning.
- It achieved a 90% detection rate on the AIxCC 2025 Final Competition C/C++ dataset.
- In real-world deployment, it discovered 29 zero-day vulnerabilities across 12 open-source projects, all confirmed and fixed.