All You Need Is A Fuzzing Brain: Our Key Achievements in DARPA’s AI Cyber Challenge

Published: September 2025

Introduction

We’re thrilled to share the technical details behind our team’s success in DARPA’s Artificial Intelligence Cyber Challenge (AIxCC). Our team, All You Need Is A Fuzzing Brain, achieved 4th place among seven finalists, autonomously discovering 28 security vulnerabilities including 6 previously unknown zero-days in real-world open-source projects.

This blog post provides an in-depth look at our LLM-powered Cyber Reasoning System (CRS) that made this achievement possible. Our complete system is open source on GitHub.

The Challenge: Autonomous Vulnerability Discovery

DARPA’s AIxCC challenged teams to build systems that could:

  1. Generate Proofs-of-Vulnerability (POVs): Create inputs that trigger sanitizer errors in fuzzing harnesses
  2. Produce Patches: Generate diff files that fix vulnerabilities while preserving functionality
  3. Operate Autonomously: Work across three challenge modes (Delta-Scan, Full-Scan, Report-Based)

The competition featured real-world C/C++ and Java projects, with scoring based on accuracy and speed.

Our Approach: The FuzzingBrain Architecture

System Overview

FuzzingBrain consists of four core services running in parallel across ~100 VMs:

  • CRS Web Service: Central coordinator handling task decomposition and fuzzer distribution
  • Static Analysis Service: Provides function metadata, reachability analysis, and call paths
  • Worker Services: Execute 23 parallel LLM-based strategies for POV/patch generation
  • Submission Service: Manages deduplication, SARIF validation, and bundle creation

Massive Parallelization

Our key insight was that parallelization is everything. We deployed:

  • ~100 virtual machines with 32-192 cores each
  • 100-10,000 concurrent threads per VM
  • 23 distinct LLM-based strategies running simultaneously
  • Multiple fuzzing harnesses processed concurrently

This architecture enabled us to discover most vulnerabilities within the first 30 minutes of analysis.

Technical Deep Dive: LLM-Powered Strategies

POV Generation: 10 Specialized Strategies

Our POV generation employs iterative, dialogue-based interaction with LLMs:

Base Strategy (xs0_delta)

System Prompt → User Message (diff + harness) → LLM Response → 
Python Execution → Fuzzer Test → Feedback Loop

Advanced Enhancements

Multi-Input Generation: Generate 5 test cases per iteration instead of 1

Coverage-Guided Feedback: For failed attempts, we provide:

  • Executed functions with ±3 lines of context
  • Branch decisions and control flow information
  • LCOV-style reports for C/C++, JaCoCo for Java

Vulnerability Category Prompting: Targeted prompts for specific CWE classes:

  • C/C++: Buffer overflows, use-after-free, integer overflows (10 categories)
  • Java: Deserialization, injection, XXE processing (15 categories)

Patching: 13 Sophisticated Strategies

Our patching follows a rigorous 7-step workflow:

  1. Target Function Identification: LLM-based analysis to find vulnerable functions
  2. Metadata Extraction: Complete source code and context retrieval
  3. Patch Generation: LLM produces revised function implementations
  4. Function Rewrite: Replace original with LLM-generated content
  5. Diff Creation: Generate standard .diff files
  6. Validation: Compilation + POV testing + functionality tests
  7. Iterative Refinement: Structured feedback for failed attempts

Novel XPatch Strategy

Our XPatch approach generates patches even without POVs:

  • Delta-scan: Analyze all modified functions from commit diffs
  • Full-scan: LLM scoring of fuzzer-reachable functions (1-10 scale)
  • Validation: 60-second LibFuzzer runs to detect new crashes

Multi-Model Resilience

We implemented a robust fallback mechanism across frontier LLMs:

  • claude-3.7, chatgpt-latest, claude-opus-4, o3, gemini-2.5-pro
  • Up to 5 iterations per model before automatic fallback
  • Leverages complementary strengths of different models

When one model fails, the system seamlessly transitions to the next, maximizing success rates.

Static Analysis: Custom Toolchains

C/C++ Analysis Pipeline

  • LLVM + SVF + Bear: Generate bitcode → construct call graphs → compute reachability
  • Engineering Challenges: Missing headers, large bitcode files (>50MB)
  • Solutions: Parallel analysis, trimming, 10-minute timeouts
  • Performance: >95% success rate on curl, dropbear, sqlite3

Java Analysis with CodeQL

  • Custom Queries: Reachable functions and call path computation
  • Batch Processing: 1,000 fuzzer-target pairs per batch
  • Isolated Execution: Cloned databases prevent race conditions
  • Speed: <5 minutes for Apache Zookeeper, Tika, Commons-Compress

Competition Results: By the Numbers

Final Round Performance

  • 28 total vulnerabilities discovered
  • 6 zero-day vulnerabilities (previously unknown)
  • 14 successful patches with functionality preservation
  • 4th place out of 7 finalist teams
  • 60 total challenges in the final round

Key Performance Insights

Speed: Most vulnerabilities found within 30 minutes Effectiveness: LLM strategies discovered nearly all POVs vs. traditional fuzzing Scalability: Successfully handled 50+ fuzzers per project Accuracy: Maintained high precision through rigorous validation

Lessons Learned & Optimizations

Resource Management

  • Sanitizer Selection: Prioritized AddressSanitizer over Memory/UndefinedBehavior
  • Time Budgeting: 60-minute caps on LLM fuzzing, 45 minutes if POVs exist
  • Patch Submission Limits: Max 5 POV-based + 3 XPatches per vulnerability

API Cost Control

We exhausted OpenAI credits in one round due to workers assigned to unreachable code. Our solution:

  • Static analysis preprocessing to identify reachable targets
  • Early termination for impossible scenarios
  • Resource allocation based on reachability analysis

Open Source Contributions

FuzzingBrain CRS

Complete system implementation with all 23 strategies: github.com/o2lab/afc-crs-all-you-need-is-a-fuzzing-brain

LLM Vulnerability Detection Leaderboard

Benchmarking framework for comparing LLMs on security tasks: fuzzingbrain.github.io/FuzzingBrain-Leaderboard

Research Team

Our interdisciplinary team spanned three institutions:

Texas A&M University: Jeff Huang (PI), Ze Sheng, Qingxiao Xu, Jianwei Huang, Matthew Woodcock, Guofei Gu

City University of Hong Kong: Heqing Huang

Imperial College London: Alastair F. Donaldson

Looking Forward

The success of FuzzingBrain demonstrates the immense potential of LLM-powered security analysis. Key areas for future research:

  • Model Specialization: Training domain-specific models for vulnerability detection
  • Hybrid Approaches: Better integration of symbolic execution and dynamic analysis
  • Scalability: Extending to larger codebases and more programming languages
  • Real-time Defense: Applying techniques to production security monitoring

Conclusion

Our 4th place finish in DARPA’s AIxCC represents a significant milestone in autonomous vulnerability discovery. By combining massive parallelization, sophisticated LLM strategies, and robust engineering, we created a system capable of finding critical security flaws at unprecedented scale and speed.

The cybersecurity landscape is evolving rapidly, and we believe LLM-powered analysis will play an increasingly crucial role in defending against emerging threats. We’re excited to continue pushing the boundaries of what’s possible in automated security research.

Want to learn more? Check out our complete technical paper and explore our open source implementation.

Questions or collaboration opportunities? Reach out to our team at [contact information].