FuzzingBrain V2
#91 en Frameworks d'agentsUnknown · v2 · depuis 2026-05-20 · 2× · vu le 29 juin 2026
FuzzingBrain V2 is a multi-agent LLM system developed at Texas A&M University (authors: Ze Sheng, Zhicheng Chen, Qingxiao Xu, Kewen Zhu, Jeff Huang) for fully automated software vulnerability discovery and reproduction, published as an arXiv preprint on May 20, 2026. The system is built on Google's OSS-Fuzz infrastructure and uses the Model Context Protocol (MCP) as the communication standard between specialized sub-agents. In benchmarks on the AIxCC 2025 dataset, it achieved a 90% detection rate (36 of 40 vulnerabilities); in real-world deployment, 29 zero-day vulnerabilities were discovered across 12 open-source projects and confirmed by maintainers. The system currently focuses on C/C++ with limited Java support via Jazzer.
Fonctionnalités
| Autonomy Level | Fully automated, no human intervention required; autonomously generates PoC inputs and vulnerability reports |
| Deployment Duration (Long-Running) | Configurable time limits: 120 minutes (delta scan) and 240 minutes (full scan) per challenge; first finding result in under 5 minutes (per fuzzingbrain.github.io) |
| Channels | Distributed worker architecture with Redis message queue and MongoDB; Analyzer Server via RPC |
| Modalities | Code analysis (text/source code), fuzzing feedback (libFuzzer with ASan/MSan/UBSan) |
| Platform | C/C++ software projects, integrated with Google's OSS-Fuzz (over 1,000 open-source projects) |
| Price Tier | Open source (Apache 2.0 license); evaluation budget in paper: $150 (delta scan) or $400 (full scan) per challenge; $200 per project in real-world deployment (API costs for LLM calls) |
| Release Date | arXiv paper published May 2026 |
| Tool/MCP Integrations | MCP-based static and dynamic analysis tools (call graphs, function extraction, reachability analysis) via BaseAgent class |