Google Continuous Fuzzing

Google Continuous Fuzzing⁠‍⁠​‌​‌​​‌‌‍​‌​​‌​‌‌‍​​‌‌​​​‌‍​‌​​‌‌​​‍​​​​​​​‌‍‌​​‌‌​‌​‍‌​​​​​​​‍‌‌​​‌‌‌‌‍‌‌​​​‌​​‍‌‌‌‌‌‌​‌‍‌‌​‌​​​​‍​‌​‌‌‌‌‌‍​‌​​‌​‌‌‍​‌‌​‌​​‌‍‌​‌​‌‌‌​‍​​‌​‌​​​‍‌‌‌​‌​‌‌‍‌​​​‌‌‌​‍​‌​​​‌​‌‍‌​​‌‌‌​‌‍‌​‌​‌​​​‍​​​​‌​‌​‍‌​​​​​‌​⁠‍⁠

Overview

Google's continuous fuzzing infrastructure (OSS-Fuzz + ClusterFuzz) has found over 10,000 bugs in 1,000+ open source projects, including critical security vulnerabilities like Heartbleed-class bugs. This technique turns fuzzing from a one-time activity into a continuous quality gate.

References

• Paper: "OSS-Fuzz - Google's continuous fuzzing service for open source software" (USENIX Security '17)
• Documentation: https://google.github.io/oss-fuzz/
• ClusterFuzz: https://google.github.io/clusterfuzz/

Core Philosophy

"Fuzzing should be continuous, not a one-time event."

"Every bug found by fuzzing is a bug not found by attackers."

Fuzzing is most effective when it runs continuously against the latest code, with automatic bug reporting and regression tracking.

Key Concepts

Coverage-Guided Fuzzing

Traditional Fuzzing:     Random input generation
Coverage-Guided Fuzzing: Inputs that increase code coverage are kept

Corpus → Mutate → Execute → Measure Coverage → Keep interesting inputs
   ↑                                                      |
   └──────────────────────────────────────────────────────┘

The Fuzzing Pipeline

1. Build: Compile with sanitizers (ASan, MSan, UBSan)
2. Fuzz: Run fuzzers continuously on cluster
3. Triage: Automatically deduplicate and file bugs
4. Reproduce: Generate minimal reproducer
5. Verify: Confirm fix eliminates the bug
6. Regress: Add reproducer to regression corpus

When Implementing

Always

• Use sanitizers (AddressSanitizer, MemorySanitizer, UndefinedBehaviorSanitizer)
• Build seed corpus from existing tests and real inputs
• Integrate fuzzing into CI/CD pipeline
• Track coverage metrics over time
• Minimize reproducers for easier debugging
• Keep regression tests for all found bugs

Never

• Fuzz only once and declare victory
• Ignore crashes in dependencies
• Skip sanitizers to "improve performance"
• Discard valuable corpus data
• Treat fuzzing as separate from testing

Prefer

• LibFuzzer/AFL++ over basic random testing
• Structure-aware fuzzing for complex formats
• Continuous fuzzing over periodic runs
• Automated triage over manual analysis
• Coverage metrics over time-based metrics

Implementation Patterns

Basic Fuzz Target (C/C++)

// fuzz_target.cc
// A fuzz target is a function that takes arbitrary bytes

#include <stdint.h>
#include <stddef.h>

// Your library headers
#include "parser.h"

extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
    // Call the function under test with fuzzer-provided data
    parse_input(data, size);
    
    // Return 0 - non-zero return values are reserved
    return 0;
}

// Build with:
// clang++ -g -fsanitize=address,fuzzer fuzz_target.cc parser.cc -o fuzzer
// Run with:
// ./fuzzer corpus_dir/

Fuzz Target with Structure

// Structure-aware fuzzing for better coverage

#include <stdint.h>
#include <stddef.h>
#include <string.h>

// Fuzz a function expecting a specific structure
struct Header {
    uint32_t magic;
    uint32_t version;
    uint32_t length;
    uint8_t  flags;
};

extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
    // Need at least header size
    if (size < sizeof(Header)) {
        return 0;
    }
    
    Header header;
    memcpy(&header, data, sizeof(Header));
    
    // Constrain to valid magic (helps fuzzer find deeper paths)
    if (header.magic != 0xDEADBEEF) {
        return 0;
    }
    
    // Constrain length to available data
    size_t payload_size = size - sizeof(Header);
    if (header.length > payload_size) {
        header.length = payload_size;
    }
    
    const uint8_t *payload = data + sizeof(Header);
    
    // Now fuzz with valid-looking input
    process_packet(&header, payload, header.length);
    
    return 0;
}

Python Fuzzing with Atheris

#!/usr/bin/env python3
# fuzz_json_parser.py

import atheris
import sys

# Import the module to fuzz
import json

def test_one_input(data):
    """Fuzz target: called with random bytes"""
    fdp = atheris.FuzzedDataProvider(data)
    
    # Convert bytes to string for JSON parsing
    json_str = fdp.ConsumeUnicodeNoSurrogates(
        fdp.ConsumeIntInRange(0, 1024)
    )
    
    try:
        # This should never crash, only raise ValueError
        json.loads(json_str)
    except (json.JSONDecodeError, ValueError):
        pass  # Expected for invalid input
    except Exception as e:
        # Unexpected exception = potential bug
        raise

def main():
    atheris.Setup(sys.argv, test_one_input)
    atheris.Fuzz()

if __name__ == "__main__":
    main()

# Run with:
# python fuzz_json_parser.py corpus_dir/ -max_len=1024

Go Fuzzing (Native)

// fuzz_test.go
// Go 1.18+ has built-in fuzzing support

package parser

import (
    "testing"
)

func FuzzParseInput(f *testing.F) {
    // Seed corpus with known inputs
    f.Add([]byte("valid input"))
    f.Add([]byte("{\"key\": \"value\"}"))
    f.Add([]byte(""))
    
    f.Fuzz(func(t *testing.T, data []byte) {
        // Call function under test
        result, err := ParseInput(data)
        
        if err != nil {
            // Errors are fine, panics are not
            return
        }
        
        // Optionally verify invariants
        if result != nil && result.Length < 0 {
            t.Errorf("negative length: %d", result.Length)
        }
    })
}

// Run with:
// go test -fuzz=FuzzParseInput -fuzztime=60s

OSS-Fuzz Integration

# Dockerfile for OSS-Fuzz integration

FROM gcr.io/oss-fuzz-base/base-builder

RUN apt-get update && apt-get install -y \
    make \
    autoconf \
    automake \
    libtool

# Clone your project
RUN git clone --depth 1 https://github.com/your/project.git

WORKDIR project
COPY build.sh $SRC/

#!/bin/bash
# build.sh - OSS-Fuzz build script

# Build the library with fuzzing instrumentation
./configure
make clean
make -j$(nproc) CC="$CC" CXX="$CXX" CFLAGS="$CFLAGS" CXXFLAGS="$CXXFLAGS"

# Build fuzz targets
$CXX $CXXFLAGS $LIB_FUZZING_ENGINE \
    fuzz_target.cc -o $OUT/fuzz_target \
    -I. libproject.a

# Copy seed corpus
zip -j $OUT/fuzz_target_seed_corpus.zip seeds/*

# Copy dictionary if available
cp project.dict $OUT/fuzz_target.dict

Corpus Management

# corpus_manager.py
# Manage and minimize fuzzing corpus

import subprocess
import hashlib
import os
from pathlib import Path

class CorpusManager:
    def __init__(self, corpus_dir: str):
        self.corpus_dir = Path(corpus_dir)
        self.corpus_dir.mkdir(exist_ok=True)
    
    def add(self, data: bytes) -> str:
        """Add input to corpus with content-based filename"""
        hash_name = hashlib.sha256(data).hexdigest()[:16]
        path = self.corpus_dir / hash_name
        
        if not path.exists():
            path.write_bytes(data)
        
        return str(path)
    
    def minimize(self, fuzzer_binary: str) -> int:
        """Minimize corpus using fuzzer's merge feature"""
        minimized_dir = self.corpus_dir.parent / "corpus_minimized"
        minimized_dir.mkdir(exist_ok=True)
        
        # LibFuzzer merge minimizes corpus
        result = subprocess.run([
            fuzzer_binary,
            "-merge=1",
            str(minimized_dir),
            str(self.corpus_dir)
        ], capture_output=True)
        
        return len(list(minimized_dir.iterdir()))
    
    def get_coverage_report(self, fuzzer_binary: str) -> dict:
        """Generate coverage report for corpus"""
        # Run with coverage instrumentation
        result = subprocess.run([
            fuzzer_binary,
            "-runs=0",  # Don't generate new inputs
            str(self.corpus_dir)
        ], capture_output=True, text=True)
        
        # Parse coverage from output
        # (actual implementation depends on sanitizer output format)
        return {"corpus_size": len(list(self.corpus_dir.iterdir()))}

CI/CD Integration

# .github/workflows/fuzz.yml
name: Continuous Fuzzing

on:
  push:
    branches: [main]
  schedule:
    - cron: '0 0 * * *'  # Daily

jobs:
  fuzz:
    runs-on: ubuntu-latest
    
    steps:
      - uses: actions/checkout@v3
      
      - name: Build fuzzer
        run: |
          clang++ -g -O1 \
            -fsanitize=address,fuzzer \
            -fno-omit-frame-pointer \
            fuzz_target.cc -o fuzzer
      
      - name: Download corpus
        uses: actions/cache@v3
        with:
          path: corpus
          key: fuzz-corpus-${{ github.sha }}
          restore-keys: fuzz-corpus-
      
      - name: Run fuzzer
        run: |
          mkdir -p corpus
          timeout 600 ./fuzzer corpus/ -max_total_time=600 || true
      
      - name: Upload crash artifacts
        if: always()
        uses: actions/upload-artifact@v3
        with:
          name: crashes
          path: crash-*
          if-no-files-found: ignore
      
      - name: Check for crashes
        run: |
          if ls crash-* 1> /dev/null 2>&1; then
            echo "Crashes found!"
            exit 1
          fi

Mental Model

Google's fuzzing approach asks:

1. Is this running continuously? One-time fuzzing misses regression bugs
2. Are sanitizers enabled? Crashes without sanitizers miss real bugs
3. Is the corpus growing? Coverage should increase over time
4. Are bugs being tracked? Automatic filing and deduplication
5. Are fixes verified? Reproducers become regression tests

Signature Moves

• Coverage-guided mutation (LibFuzzer, AFL++)
• Sanitizer builds (ASan, MSan, UBSan, TSan)
• Automatic corpus management and minimization
• CI/CD integration for every commit
• Regression corpus from found bugs
• Structure-aware fuzzing for protocols