cufuzz-artifacts/
├── src/                        # Core cuFuzz components
│   ├── cufuzz_cov_nvbit/       # NVBit-based device-side coverage collection tool
│   └── cufuzz_sand/            # Sanitizer wrappers for SAND integration
├── targets/                    # Example fuzzing targets
│   └── sampleApp/              # Simple CUDA app with intentional bug for testing
├── scripts/                    # Evaluation and analysis scripts
├── Tools/                      # External dependencies
│   ├── AFLplusplus/            # AFL++ fuzzer (git submodule)
│   └── AFLplusplus.patch       # cuFuzz patches for AFL++
├── third-party-licenses/       # Third-party license files
│   ├── LICENSE_from_aflplusplus  # AFL++ Apache 2.0 license
│   └── LICENSE_from_nvbit        # NVBit NVIDIA EULA
├── images/                     # Documentation images
├── build.sh                    # Automated build script
├── verify_build.sh             # Quick verification test
├── Dockerfile                  # Docker container definition
├── LICENSE                     # Apache License 2.0
└── CONTRIBUTING.md             # Contribution guidelines and DCO

tar -xzvf cufuzz-artifacts.tar.gz
cd cufuzz-artifacts

# Clone AFL++ (required dependency)
git clone https://github.com/AFLplusplus/AFLplusplus.git Tools/AFLplusplus
cd Tools/AFLplusplus
git checkout 9cac7ced05eb9f36c1d0b02ad594b3b09cd3938b
cd ../..

sudo docker build --build-arg GPU_ARCH=<your_arch> -t cufuzz .

sudo docker build --build-arg GPU_ARCH=sm_86 -t cufuzz .

sudo docker run --rm --gpus all -it -v /:/my_workspace cufuzz bash

root@container:~/cufuzz# ./verify_build.sh

apt-get update && apt-get install -y build-essential python3-dev automake cmake git flex \
    bison libglib2.0-dev libpixman-1-dev python3-setuptools cargo libgtk-3-dev lld llvm llvm-dev \
    clang ninja-build cpio libcapstone-dev wget curl python3-pip vim less libxxhash-dev bc zlib1g-dev

export GPU_ARCH=sm_86  # Change to match your GPU

cd Tools/AFLplusplus 
patch -N -p1 < ../AFLplusplus.patch 
export CXX=/usr/bin/clang++-14 
export CC=/usr/bin/clang-14
make -j8 &> build.log

mkdir -p Tools/NVBit
wget https://github.com/NVlabs/NVBit/releases/download/v1.7.5/nvbit-Linux-x86_64-1.7.5.tar.bz2
tar -xvf nvbit-Linux-x86_64-1.7.5.tar.bz2 
mv nvbit_release_x86_64/* Tools/NVBit/
rm -rf nvbit_release_x86_64 nvbit-Linux-x86_64-1.7.5.tar.bz2

cd src/cufuzz_cov_nvbit/
export GPU_ARCH=sm_86  # Adjust for your GPU
ARCH=$GPU_ARCH make 

cd src/cufuzz_sand
AFL_SAN_NO_INST=1 ../../Tools/AFLplusplus/afl-clang-fast -O2 wrapper_san.c -o wrapper_memcheck.out 
AFL_SAN_NO_INST=1 ../../Tools/AFLplusplus/afl-clang-fast -DSAN_MODE_INIT -O2 wrapper_san.c -o wrapper_initcheck.out 
AFL_SAN_NO_INST=1 ../../Tools/AFLplusplus/afl-clang-fast -DSAN_MODE_RACE -O2 wrapper_san.c -o wrapper_racecheck.out
AFL_SAN_NO_INST=1 ../../Tools/AFLplusplus/afl-clang-fast -DSAN_MODE_ASAN -O2 wrapper_san.c -o wrapper_asan.out

CUFUZZ_MAP_SIZE=65536 AFL_SKIP_CPUFREQ=1 AFL_PRELOAD=/PATH/TO/cufuzz_cov.so \
    ./Tools/AFLplusplus/afl-fuzz -x sample.dict -i input_samples/ -o output_dir/ \
    -t 1000000 ./cuda_app.out @@

cd src/cufuzz_sand

# Build sanitizer wrappers
AFL_SAN_NO_INST=1 ../../Tools/AFLplusplus/afl-clang-fast -O2 wrapper_san.c -o wrapper_memcheck.out 
AFL_SAN_NO_INST=1 ../../Tools/AFLplusplus/afl-clang-fast -DSAN_MODE_INIT -O2 wrapper_san.c -o wrapper_initcheck.out 
AFL_SAN_NO_INST=1 ../../Tools/AFLplusplus/afl-clang-fast -DSAN_MODE_RACE -O2 wrapper_san.c -o wrapper_racecheck.out
AFL_SAN_NO_INST=1 ../../Tools/AFLplusplus/afl-clang-fast -DSAN_MODE_ASAN -O2 wrapper_san.c -o wrapper_asan.out

cd ../../targets/sampleApp/

export PATH=/usr/local/cuda/bin/:$PATH
export GPU_ARCH=sm_86  # Adjust for your GPU

# Build vanilla version (for sanitizer)
nvcc sampleApp.cu -I/usr/local/cuda/include/ -O2 --ptxas-options "-v" \
    --gpu-architecture=$GPU_ARCH -o sampleApp-vanilla.out

# Build instrumented version (for fuzzing)
nvcc sampleApp.cu -I/usr/local/cuda/include/ -O2 --ptxas-options "-v" \
    --gpu-architecture=$GPU_ARCH --compiler-bindir ../../Tools/AFLplusplus/afl-clang-fast++ \
    -o sampleApp.out

# Run cuFuzz
ORIGINAL_APP=./sampleApp-vanilla.out \
SANITIZER_PATH=/usr/local/cuda/bin/compute-sanitizer \
SANITIZER_ARG="--tool=memcheck --report-api-errors=no --error-exitcode 99" \
SANITIZER_ARG_RACE="--tool=racecheck --report-api-errors=no --error-exitcode 99" \
SANITIZER_ARG_INIT="--tool=initcheck --report-api-errors=no --error-exitcode 99" \
CUFUZZ_MAP_SIZE=65536 \
AFL_SKIP_CPUFREQ=1 \
AFL_PRELOAD=../../src/cufuzz_cov_nvbit/cufuzz_cov.so \
../../Tools/AFLplusplus/afl-fuzz -x sample.dict -i in/ -o out/ \
    -w ../../src/cufuzz_sand/wrapper_memcheck.out \
    -w ../../src/cufuzz_sand/wrapper_racecheck.out \
    -w ../../src/cufuzz_sand/wrapper_initcheck.out \
    -t 1000000 ./sampleApp.out @@

cd targets/sampleApp/

export PATH=/usr/local/cuda/bin/:$PATH
export GPU_ARCH=sm_86  # Adjust for your GPU

# Build persistent mode binary
nvcc sampleApp_persistent.cu -I/usr/local/cuda/include/ -O2 --ptxas-options "-v" \
    --gpu-architecture=$GPU_ARCH --compiler-bindir ../../Tools/AFLplusplus/afl-clang-fast++ \
    -o sampleApp_persistent.out

# Run cuFuzz in persistent mode
COV_PERSISTENT=1 \
CUFUZZ_MAP_SIZE=65536 \
AFL_SKIP_CPUFREQ=1 \
AFL_PRELOAD=../../src/cufuzz_cov_nvbit/cufuzz_cov.so \
./../../Tools/AFLplusplus/afl-fuzz -x sample.dict -i in/ -o out/ \
    -t 1000000 ./sampleApp_persistent.out @@

@article{cufuzz2026,
  title={Hunting CUDA Bugs at Scale with cuFuzz},
  author={Mohamed Tarek Ibn Ziad and Christos Kozyrakis},
  journal={Proceedings of the ACM on Programming Languages},
  volume={10},
  number={OOPSLA1},
  article={123},
  month={4},
  year={2026},
  doi={10.1145/3798231}
}

Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0