Finding Vulnerabilities Through Fuzzing

Overview

Fuzzing can feel a bit front‑loaded: you may spend time wiring harnesses and running campaigns without immediately finding exciting new bugs, especially on hardened or well‑tested targets. That’s normal, and it's one reason the next week on patch diffing often feels more directly "practical" — many companies already run large fuzzing setups and need people who can understand and exploit the bugs those systems uncover. Still, working through this week is important: it teaches you how fuzzers actually discover real vulnerabilities, so when you later triage crashes or study patches, you'll have a solid intuition for how those bugs were found and how to reproduce them.

Prerequisites

Before starting this week, ensure you have:

• A Linux virtual machine (Ubuntu 24.04 recommended) with at least 8GB RAM and 8 cpu cores

• Basic understanding of C/C++ programming

• Familiarity with command-line tools and debugging (GDB basics)

• Understanding of memory corruption vulnerabilities (from Week 1)

Day 1: Introduction to Fuzzing

• Goal: Understand the fundamentals of fuzzing and get hands-on experience with AFL++.

• Activities:

  • Reading: "Fuzzing for Software Security Testing and Quality Assurance" by Ari Takanen(From 1.3.2 to 1.3.8 and 2.4.1 to 2.7.5).

  • Online Resource:

    • Fuzzing Book by Andreas Zeller - Read "Introduction" and "Fuzzing Basics."

    • AFL++ Documentation - Follow the quick start guide.

    • Interactive Module to Learn Fuzzing

  • Real-World Context:

    • Google OSS-Fuzz: Finding 36,000+ bugs across 1,000+ projects

    • AFL Success Stories - Real vulnerabilities found by AFL

  • Exercise:

    • Set up a Linux virtual machine (VM) with the necessary tools installed, including compilers and debuggers

    • Run AFL++ on a C program

    • If possible, use or write a small C program that contains a simple version of one of the Week 1 vulnerability classes (for example, a stack buffer overflow or integer overflow) so you can see fuzzing rediscover it.

# Setting up AFL++

# Install build dependencies
sudo apt update
sudo apt install -y build-essential gcc-13-plugin-dev cpio python3-dev libcapstone-dev \
    pkg-config libglib2.0-dev libpixman-1-dev automake autoconf python3-pip \
    ninja-build cmake git wget python3.12-venv meson

# Install LLVM (check latest version at https://apt.llvm.org/)
wget https://apt.llvm.org/llvm.sh
chmod +x llvm.sh
sudo ./llvm.sh 19 all

# Verify LLVM installation
clang-19 --version
llvm-config-19 --version

# Install Rust (required for some AFL++ components)
curl --proto '=https' --tlsv1.2 -sSf "https://sh.rustup.rs" | sh
source ~/.cargo/env

# Build and install AFL++
mkdir -p ~/soft && cd ~/soft
git clone --depth 1 https://github.com/AFLplusplus/AFLplusplus.git
cd AFLplusplus
# NOTE: unicorn support might fail(you need to add the env or run ./build_unicorn_support.py and fix issues yourself)
make distrib
sudo make install

# Verify installation
which afl-fuzz
afl-fuzz --version

# Phase 1: Simple crash example
cd ~/ && mkdir -p tuts && cd tuts
git clone --branch main --depth 1 https://github.com/alex-maleno/Fuzzing-Module.git
cd Fuzzing-Module/exercise1 && mkdir -p build && cd build

# Compile with AFL++ instrumentation
CC=/usr/local/bin/afl-clang-fast CXX=/usr/local/bin/afl-clang-fast++ cmake ..
make

# Create seed inputs
cd .. && mkdir -p seeds && cd seeds
for i in {0..4}; do
    dd if=/dev/urandom of=seed_$i bs=64 count=10 2>/dev/null
done

# Run AFL++ fuzzer
cd ../build
echo core | sudo tee /proc/sys/kernel/core_pattern
afl-fuzz -i ../seeds/ -o out -m none -d -- ./simple_crash

# Expected output: AFL++ interface showing coverage, crashes, etc.
# Look for crashes in out/crashes/ directory

# Phase 2: Medium complexity example
cd ~/tuts/Fuzzing-Module/exercise2 && mkdir -p build && cd build
CC=/usr/local/bin/afl-clang-lto CXX=/usr/local/bin/afl-clang-lto++ cmake ..
make

cd .. && mkdir -p seeds && cd seeds
for i in {0..4}; do
    dd if=/dev/urandom of=seed_$i bs=64 count=10 2>/dev/null
done

cd ../build
afl-fuzz -i ../seeds/ -o out -m none -d -- ./medium

Success Criteria:

• AFL++ compiles and installs without errors

• Both fuzzing sessions start successfully

• You can see the AFL++ status screen showing paths found, crashes, etc.

• Check out/crashes/ directory for any discovered crashes

Troubleshooting:

• If afl-clang-fast not found: Check /usr/local/bin/ is in PATH

• If compilation fails: Ensure LLVM 19 is properly installed (clang-19 --version)

• If fuzzer doesn't start: Check CPU scaling governor (echo performance | sudo tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor)

Real-World Impact: AFL++ Finding CVE-2024-47606 (GStreamer)

Background: AFL++ and similar fuzzers are actively used to find vulnerabilities in production software. Let's examine a real case from Week 1.

Case Study - CVE-2024-47606 (GStreamer Signed-to-Unsigned Integer Underflow):

• Discovery Method: Continuous fuzzing campaigns by security researchers using AFL++ on media parsers

• The Bug: GStreamer's qtdemux_parse_theora_extension had a signed integer underflow that became massive unsigned value

• Attack Surface: MP4/MOV files processed automatically by browsers, media players, messaging apps

• Fuzzing Approach:

  1. Target: GStreamer's QuickTime demuxer (qtdemux)

  2. Seed corpus: Valid MP4 files from public datasets

  3. Instrumentation: Compiled with AFL++ and AddressSanitizer

  4. Mutation strategy: Structure-aware (understanding MP4 atoms)

  5. Result: Heap buffer overflow crash after ~48 hours of fuzzing

Why Fuzzing Found It:

• Rare Input Combination: Required specific Theora extension size values that underflow

• Static Analysis Limitation: Signed-to-unsigned conversion buried in complex parsing logic

• Code Review Miss: Integer arithmetic looked correct without considering negative values

• Automated Testing Gap: Unit tests didn't cover malformed Theora extensions

The Discovery Process:

# 1) Generate a structured MP4 seed corpus (GitHub Security Lab generator)
cd ~/tuts && git clone --depth 1 https://github.com/github/securitylab.git
cd ~/tuts/securitylab/Fuzzing/GStreamer
make
mkdir -p corpus/mp4
./generator -o corpus/mp4

# 2) Build a vulnerable GStreamer (< 1.24.10) with AFL++ + ASan
cd ~/tuts
git clone --branch 1.24.9 --depth 1 https://gitlab.freedesktop.org/gstreamer/gstreamer.git
cd gstreamer
export CC=afl-clang-fast
export CXX=afl-clang-fast++
export CFLAGS="-O1 -g"
export CXXFLAGS="-O1 -g"
sudo apt-get install -y flex bison
# NOTE: this might take a while so you can just build parts of it, not all
meson setup build-afl --buildtype=debug -Db_sanitize=address
ninja -C build-afl -j"$(nproc)"

# 3) Fuzz the QuickTime demuxer pipeline with AFL++
mkdir -p findings
# NOTE: you can fuzz other binaries as well to find bugs
echo core | sudo tee /proc/sys/kernel/core_pattern
afl-fuzz -i ~/tuts/securitylab/Fuzzing/GStreamer/corpus/mp4 \
         -o findings -m none -- \
         ./build-afl/subprojects/gstreamer/tools/gst-launch-1.0 \
         filesrc location=@@ ! qtdemux ! fakesink

# Typical outcome after hours of fuzzing:
#   - ASan crash inside qtdemux_parse_theora_extension()
#   - heap-buffer-overflow in gst_buffer_fill() when copying attacker-controlled data
# Root cause (CVE-2024-47606 / GHSL-2024-166, fixed in 1.24.10):
#   - 32-bit signed 'size' underflows → huge unsigned value
#   - _sysmem_new_block() overflows when adding alignment/header → tiny (0x89-byte) allocation
#   - memcpy() writes the huge size, corrupting GstMapInfo and allocator function pointers

Key Insight: Fuzzing excels at finding edge cases in complex parsers that humans would never manually test. The combination of:

• Coverage-guided mutation (AFL++ exploring new code paths)

• AddressSanitizer (detecting memory corruption immediately)

• Persistent fuzzing (running for days/weeks)

...makes it more effective than manual testing for this vulnerability class.

Key Takeaways

1. Fuzzing finds real vulnerabilities: Not just theoretical crashes, but exploitable bugs in production software

2. Coverage-guided fuzzing is powerful: AFL++ intelligently explores code paths rather than random mutation

3. Sanitizers are essential: ASAN, UBSAN turn subtle bugs into immediate crashes

4. Time matters: Many bugs require hours/days of fuzzing to discover

5. Seed corpus quality affects results: Starting with valid inputs helps reach deeper code paths

Discussion Questions

1. Why did fuzzing find CVE-2024-47606 when code review and unit testing didn't?

2. What advantages does coverage-guided fuzzing have over purely random fuzzing?

3. How do sanitizers (ASAN, UBSAN) enhance fuzzing effectiveness?

4. What types of vulnerabilities are fuzzing best suited to find? What types does it miss?

5. How can seed corpus selection impact fuzzing effectiveness?

Day 2: Continue Fuzzing with AFL++

• Goal: Understand and apply advanced fuzzing techniques.

• Activities:

  • Reading: Continue with "Fuzzing for Software Security Testing and Quality Assurance" (From 3.3 to 3.9.8).

  • Real-World Examples:

    • AFL++ finds CVE-2020-9385 in ZINT Barcode Generator - Stack buffer overflow discovered through fuzzing

    • AFL++ Fuzzing in Depth - How to effectively use afl++

    • Suricata IDS CVE-2019-16411 - Out-of-bounds read found via fuzzing

  • Exercise:

    • Experiment with different AFL++ options (for example, dictionary-based fuzzing, persistent mode).

    • Running AFL++ with a real-world application like a file format parser to mimic real-world scenarios.

    • Optionally, target an image or media parser so you can practice finding heap overflows and out-of-bounds reads similar to the libWebP and GStreamer bugs from Week 1.

# Fuzzing a image parser (dlib imglab)
# NOTE: you can pull older versions to guarantee vulnerable code paths
cd ~/tuts && git clone --depth 1 --branch v19.24.6 https://github.com/davisking/dlib.git
cd dlib/tools/imglab && mkdir -p build && cd build

# Configure sanitizers for better crash detection
export AFL_USE_UBSAN=1
export AFL_USE_ASAN=1
export ASAN_OPTIONS="detect_leaks=1:abort_on_error=1:allow_user_segv_handler=0:handle_abort=1:symbolize=0"

# Install dependencies
sudo apt install -y libx11-dev libavdevice-dev libavfilter-dev libavformat-dev libavcodec-dev \
libswresample-dev libswscale-dev libavutil-dev libjxl-dev libjxl-tools

# Compile with AFL++ and sanitizers
cmake -DCMAKE_C_COMPILER=afl-clang-fast \
      -DDLIB_NO_GUI_SUPPORT=0 \
      -DCMAKE_CXX_COMPILER=afl-clang-fast++ \
      -DCMAKE_CXX_FLAGS="-fsanitize=address,leak,undefined -g" \
      -DCMAKE_C_FLAGS="-fsanitize=address,leak,undefined -g" ..
make -j$(nproc)

# Prepare seed corpus
mkdir -p fuzz/image/in
cp ../../../examples/faces/testing.xml fuzz/image/in/

# TODO: try to improve the fuzzing speed using https://aflplus.plus/docs/fuzzing_in_depth/#i-improve-the-speed
# Run AFL++ in parallel mode (Master + Slave instances)
# Terminal 1: Master instance
echo core | sudo tee /proc/sys/kernel/core_pattern
afl-fuzz -i fuzz/image/in -o fuzz/image/out -M Master -- ./imglab --stats @@

# Terminal 2: Slave instance (for parallel fuzzing)
afl-fuzz -i fuzz/image/in -o fuzz/image/out -S Slave1 -- ./imglab --stats @@

# Install crash analysis tools
sudo apt install -y gdb python3-pip valgrind
wget -O ~/.gdbinit-gef.py -q https://gef.blah.cat/py
echo "source ~/.gdbinit-gef.py" >> ~/.gdbinit

# Minimize a crashing input while preserving the crashing behavior (afl-tmin)
# NOTE: there might be no crashes, either fuzz longer or go back to an older tag
CRASH=$(ls ~/tuts/dlib/tools/imglab/build/fuzz/image/out/Master/crashes/id* 2>/dev/null | head -n1)
afl-tmin -i "$CRASH" -o ~/tuts/dlib/tools/imglab/build/fuzz/image/out/Master/crashes/minimized_crash -- ./imglab --stats @@

# Cluster and triage crashes with casr-afl (from CASR tools)
# NOTE: there might be no crashes, either fuzz longer or go back to an older tag
CASR_URL="https://github.com/ispras/casr/releases/latest/download/casr-x86_64-unknown-linux-gnu.tar.xz"
INSTALL_DIR="$HOME/.local"
mkdir -p "$INSTALL_DIR"
wget -O "$INSTALL_DIR/casr-x86_64-unknown-linux-gnu.tar.xz" "$CASR_URL"
tar -xJf "$INSTALL_DIR/casr-x86_64-unknown-linux-gnu.tar.xz" -C "$INSTALL_DIR"
export PATH="$INSTALL_DIR/casr-x86_64-unknown-linux-gnu/bin:$PATH"  # provides casr-afl

# Now run casr-afl on the AFL++ output directory
casr-afl -i ~/tuts/dlib/tools/imglab/build/fuzz/image/out/Master -o ~/tuts/dlib/tools/imglab/build/fuzz/image/out/Master_casr_reports

Expected Outputs:

• AFL++ status screen showing increasing coverage

• Crashes appearing in fuzz/image/out/Master/crashes/ or fuzz/image/out/Slave1/crashes/

• AddressSanitizer reports for memory corruption bugs

What to Look For:

• Crashes with SIGSEGV or SIGABRT signals

• AddressSanitizer reports showing heap buffer overflows, use-after-free, etc.

• Unique crash signatures (different stack traces)

Troubleshooting:

• If compilation fails: Check that all dependencies are installed

• If no crashes found: Let fuzzer run longer (hours/days for real targets)

• If crashes are false positives: Review ASAN options and adjust

Real-World Campaign: Fuzzing Image Parsers

Case Study - CVE-2023-4863 (libWebP Heap Buffer Overflow):

From Week 1, you learned about this critical vulnerability. Let's understand how fuzzing could have (and did) discover similar bugs.

• The Target: libWebP image decoder, used by Chrome, Firefox, and countless applications

• Why It's Fuzzing-Friendly:

  • Pure input-to-output: takes file bytes, produces image

  • No network/filesystem dependencies

  • Deterministic execution

  • Complex parsing logic with many edge cases

Fuzzing Campaign Strategy:

# Real-world fuzzing setup for image parsers
cd ~/tuts && git clone --depth 1 --branch 1.0.0 https://chromium.googlesource.com/webm/libwebp

cd libwebp && sudo apt-get -y install gcc make autoconf automake libtool

# Compile with AFL++ and all sanitizers
export CC=afl-clang-fast
export CXX=afl-clang-fast++
export AFL_USE_ASAN=1
export AFL_USE_UBSAN=1
export CFLAGS="-fsanitize=address,undefined -g"
export CXXFLAGS="-fsanitize=address,undefined -g"

./autogen.sh
./configure
make -j$(nproc)

# Create fuzzing harness
cat > fuzz_webp.c << 'EOF'
#include <stdint.h>
#include <stdlib.h>
#include <stdio.h>
#include <webp/decode.h>
#include <webp/types.h>

int main(int argc, char **argv) {
    if (argc < 2) return 1;

    FILE *f = fopen(argv[1], "rb");
    if (!f) return 1;

    fseek(f, 0, SEEK_END);
    size_t size = ftell(f);
    fseek(f, 0, SEEK_SET);

    uint8_t *data = malloc(size);
    fread(data, 1, size, f);
    fclose(f);

    // Fuzz target: decode WebP image
    int width, height;
    uint8_t *output = WebPDecodeRGBA(data, size, &width, &height);

    if (output) free(output);
    free(data);
    return 0;
}
EOF

# Compile fuzzing harness
afl-clang-fast -I./src -o fuzz_webp fuzz_webp.c \
    -L./src/.libs -lwebp -fsanitize=address,undefined -g

# Collect seed corpus (valid WebP images)
mkdir -p ~/tuts/libwebp/seeds
# Download some WebP test images
wget -q -O ~/tuts/libwebp/seeds/test1.webp https://www.gstatic.com/webp/gallery/1.webp
wget -q -O ~/tuts/libwebp/O seeds/test2.webp https://www.gstatic.com/webp/gallery/2.webp
wget -q -O ~/tuts/libwebp/O seeds/test3.webp https://www.gstatic.com/webp/gallery/3.webp

# Run AFL++ fuzzer
export LD_LIBRARY_PATH=./src/.libs:$LD_LIBRARY_PATH
afl-fuzz -i seeds/ -o findings/ -m none -d -- ./fuzz_webp @@

# Real campaigns run for weeks. OSS-Fuzz runs 24/7.
# Expected: Crashes in findings/crashes/ directory
# Analysis: ASAN reports showing heap buffer overflows

What Fuzzing Discovered:

In the real CVE-2023-4863 case:

1. Initial crash: Heap buffer overflow in BuildHuffmanTable()

2. Root cause: Malformed Huffman coding data caused out-of-bounds write

3. ASAN output: Immediate detection of corruption with exact location

4. Exploitability: Function pointer hijack possible via heap corruption

Why This Bug Survived Testing:

• Unit tests: Covered valid WebP files, not malformed Huffman tables

• Static analysis: Complex pointer arithmetic hard to verify

• Code review: Bounds check looked correct in isolation

• Fuzzing advantage: Generated millions of mutated WebP files, including edge cases

Parallel Fuzzing for Speed:

# Real campaigns use multiple CPU cores
# Master instance
afl-fuzz -i seeds/ -o findings/ -M master -m none -- ./fuzz_webp @@

# Slave instances (in separate terminals or tmux)
for i in {1..5}; do
    afl-fuzz -i seeds/ -o findings/ -S slave$i -m none -- ./fuzz_webp @@ &
done

# Check status
afl-whatsup findings/

# Expected output:
# Master: 1234 paths, 5 crashes
# Slave1: 987 paths, 2 crashes
# Slave2: 1056 paths, 3 crashes
# ... (instances share corpus and findings)

Corpus Management and Seed Selection

Why Seed Quality Matters:

# Bad seed corpus: random bytes
dd if=/dev/urandom of=bad_seed.webp bs=1024 count=10

# Result: AFL++ spends time on invalid inputs that fail early parsing
# Coverage: Only reaches format validation code

# Good seed corpus: valid WebP files
# Result: AFL++ mutates valid structure, reaches deep parsing logic
# Coverage: Explores Huffman decoding, color space conversion, filters

Building Effective Seed Corpus:

# 1. Collect diverse valid inputs
mkdir -p corpus
# - Different sizes (small, medium, large)
# - Different features (lossy, lossless, animated)
# - Different color spaces (RGB, YUV, alpha channel)
wget -r -l1 -A webp https://www.gstatic.com/webp/gallery/ -P corpus/

# 2. Minimize corpus (remove redundant files)
afl-cmin -i corpus/ -o corpus_min/ -- ./fuzz_webp @@

# 3. Minimize individual files (shrink while preserving coverage)
mkdir -p corpus_tmin
for f in corpus_min/*; do
    afl-tmin -i "$f" -o "corpus_tmin/$(basename $f)" -- ./fuzz_webp @@
done

# Result: Smaller corpus = faster fuzzing iterations
# Original: 50 files, 5MB total
# Minimized: 15 files, 500KB total (same coverage)

Key Takeaways

1. Image parsers are prime fuzzing targets: Complex, widely-deployed, handle untrusted input

2. OSS-Fuzz prevents 0-days: Continuous fuzzing finds bugs before attackers

3. Parallel fuzzing scales linearly: 8 cores = ~8x throughput

4. Corpus quality > corpus size: Minimized, diverse seeds outperform large random corpus

5. Dictionaries accelerate discovery: Format-aware tokens reach deeper code paths faster

Discussion Questions

1. Why are image/media parsers particularly well-suited for fuzzing compared to other software?

2. How does corpus minimization improve fuzzing efficiency without losing coverage?

3. What trade-offs exist between fuzzing speed (lightweight instrumentation) and bug detection (heavy sanitizers)?

4. Why did OSS-Fuzz find bugs in libwebp that years of production use didn't reveal?

5. How can you determine if a fuzzing campaign has reached diminishing returns and should target a different component?

6. How can you improve fuzzing speed?

Day 3: Introduction to Google FuzzTest

• Goal: Understand in-process fuzzing with FuzzTest and how to turn unit tests into coverage-guided fuzzers that actually find memory corruption bugs.

• Activities:

  • Reading: Continue with "Fuzzing for Software Security Testing and Quality Assurance" (From 4.2.1 to 4.4).

  • Online Resources:

    • Google FuzzTest - Read the README and "Getting Started".

    • Property-based fuzzing vs example-based testing - Short motivation for FuzzTest.

  • Exercises:

    1. Set up FuzzTest in a small CMake project and run a trivial property-based test.

    2. Use FuzzTest + AddressSanitizer to rediscover a simple heap buffer overflow (Week 1 vulnerability class).

    3. Extend the fuzz target to cover a small parser-style function, similar to the image/format parsers from Days 1–2.

Why FuzzTest in a vulnerability-focused course?

FuzzTest is a unit-test-style, in-process fuzzing framework from Google that:

• Integrates with GoogleTest: You write TEST and FUZZ_TEST side by side in the same file.

• Uses coverage-guided fuzzing under the hood (libFuzzer-style) but hides boilerplate harness code.

• Works great for libraries and core logic (parsers, decoders, crypto helpers) where you already have unit tests.

• Is ideal for CI: The same binary can run fast deterministic tests or long-running fuzz campaigns depending on flags.

Where AFL++/Honggfuzz are great for whole programs and black-box binaries, FuzzTest shines when you have source code and want to fuzz individual C++ functions directly.

Lab 1: Set up FuzzTest and run a basic property

mkdir -p ~/tuts/first_fuzz_project && cd ~/tuts/first_fuzz_project
git clone --branch main --depth 1 https://github.com/google/fuzztest.git

cat <<EOT > CMakeLists.txt
# GoogleTest requires at least C++17
set(CMAKE_CXX_STANDARD 17)

add_subdirectory(fuzztest)

enable_testing()

include(GoogleTest)
fuzztest_setup_fuzzing_flags()
add_executable(
  first_fuzz_test
  first_fuzz_test.cc
)

link_fuzztest(first_fuzz_test)
gtest_discover_tests(first_fuzz_test)
EOT
cat <<EOT > first_fuzz_test.cc
#include "fuzztest/fuzztest.h"
#include "gtest/gtest.h"

TEST(MyTestSuite, OnePlusTwoIsTwoPlusOne) {
  EXPECT_EQ(1 + 2, 2 + 1);
}

void IntegerAdditionCommutes(int a, int b) {
  EXPECT_EQ(a + b, b + a);
}
FUZZ_TEST(MyTestSuite, IntegerAdditionCommutes);
EOT
mkdir -p build && cd build

# configure with fuzztest
cc=clang-19 cxx=clang++-19 cmake -dcmake_build_type=relwithdebug -dfuzztest_fuzzing_mode=on ..

# Build the project
cmake --build . --parallel $(nproc)

# Run the fuzz test (short sanity run)
./first_fuzz_test --fuzz=MyTestSuite.IntegerAdditionCommutes --max_total_time=10

You should see FuzzTest/libFuzzer-style statistics (executions per second, coverage, etc.). For a correct property like integer commutativity, the fuzzer should not find crashes.

Lab 2: FuzzTest to find a heap buffer overflow

Now turn FuzzTest onto a deliberately vulnerable function that mimics a classic stack / heap buffer overflow from Week 1.

cd ~/tuts/first_fuzz_project

cat <<'EOT' > first_fuzz_test.cc
#include <cstring>
#include <string>
#include "fuzztest/fuzztest.h"
#include "gtest/gtest.h"

TEST(ArithmeticSuite, OnePlusTwoIsTwoPlusOne) {
  EXPECT_EQ(1 + 2, 2 + 1);
}

void IntegerAdditionCommutes(int a, int b) {
  EXPECT_EQ(a + b, b + a);
}
FUZZ_TEST(ArithmeticSuite, IntegerAdditionCommutes);

void VulnerableHeaderCopy(const std::string& input) {
  char header[32];

  // When input.size() > 32 and ASAN is enabled, this becomes a detectable overflow.
  std::memcpy(header, input.data(), input.size());
}

FUZZ_TEST(OverflowSuite, VulnerableHeaderCopy);
EOT

cd build

# Build the project
cmake --build . --parallel $(nproc)

# Run only the overflow fuzz test to focus on the bug
./first_fuzz_test --fuzz=OverflowSuite.VulnerableHeaderCopy --max_total_time=20

Expected result: After a short time, FuzzTest should report a crash with an AddressSanitizer message similar to:

==1066==ERROR: AddressSanitizer: stack-buffer-overflow on address 0x76f8218731c0 at pc 0x60a5060193a2 bp 0x7ffc65c1fd10 sp 0x7ffc65c1f4d0

At this point you can:

• Open first_fuzz_test.cc and fix the bug by adding a length check (for example, only copying up to sizeof(header)).

• Rebuild and re-run the fuzz target to confirm the crash is gone.

This is exactly the same pattern as our AFL++ labs: fuzzer + sanitizer → crash → root cause → fix, but now entirely inside a unit test binary.

Lab 3: Fuzzing a small parser-style function

To connect FuzzTest to the real-world parsers from Days 1–2, fuzz a tiny length-prefixed parser that can easily go wrong if you mishandle integer arithmetic.

cd ~/tuts/first_fuzz_project

cat <<'EOT' >> first_fuzz_test.cc

struct Message {
  uint8_t len;
  std::string payload;
};

Message ParseMessage(const std::string& input) {
  Message m{0, ""};
  if (input.empty()) return m;

  uint8_t len = static_cast<uint8_t>(input[0]);

  // BUG -> commenting this would cause fuzzer to find vuln
  // - If len > input.size() - 1, this will either truncate or read garbage.
  // - Here we clamp len to avoid UB, but real code often forgets this check.
  if (static_cast<size_t>(len) > input.size() - 1) {
      len = static_cast<uint8_t>(input.size() - 1);
  }

  m.len = len;
  m.payload.assign(input.data() + 1, input.data() + 1 + m.len);
  return m;
}

void ParseDoesNotCrash(const std::string& input) {
  (void)ParseMessage(input);
}

void LengthFieldRespected(const std::string& input) {
  if (input.size() < 2) return;

  uint8_t claimed_len = static_cast<uint8_t>(input[0]);
  if (static_cast<size_t>(claimed_len) > input.size() - 1) return;

  Message m = ParseMessage(input);
  EXPECT_EQ(m.len, claimed_len);
  EXPECT_EQ(m.payload.size(), claimed_len);
}

FUZZ_TEST(ParserSuite, ParseDoesNotCrash);
FUZZ_TEST(ParserSuite, LengthFieldRespected);
EOT

cd build && cmake --build . --parallel $(nproc)

# Run parser fuzz tests for a short time
./first_fuzz_test --fuzz=ParserSuite.ParseDoesNotCrash --max_total_time=15
./first_fuzz_test --fuzz=ParserSuite.LengthFieldRespected --max_total_time=15

What to look for:

• If you intentionally break the length check inside ParseMessage (for example, remove the if (len > input.size() - 1) guard - 3 lines), FuzzTest + ASAN/UBSAN should quickly find crashes or undefined behavior.

• Try modifying the parser to add more fields (flags, type bytes, nested length fields) and see how FuzzTest finds edge cases you did not think about.

Key Takeaways

1. FuzzTest brings fuzzing into your unit tests: You can turn GoogleTest-style tests into coverage-guided fuzzers with FUZZ_TEST, using the same build system and test runner.

2. Sanitizers are critical: Combining FuzzTest with ASAN/UBSAN turns memory bugs (overflows, UAFs, integer issues) into immediate, reproducible crashes.

3. Great fit for parsers and core logic: Short, pure C++ functions (parsers, decoders, protocol handlers) are ideal FuzzTest targets, similar to the real-world parsers from Days 1–2.

4. Properties > examples: Expressing invariants like “never crash” or “length field matches payload” lets the fuzzer explore inputs you would never hand-write.

5. Same workflow as other fuzzers: Regardless of tool (AFL++, Honggfuzz, FuzzTest), the basic loop is still fuzz → crash → triage → exploitability → fix.

Discussion Questions

1. In what situations would you prefer FuzzTest over a process-level fuzzer like AFL++ or Honggfuzz, and why?

2. How would you go about converting an existing GoogleTest regression test into an effective FUZZ_TEST that can find new bugs, not just regressions?

3. Which vulnerability classes from Week 1 (e.g., buffer overflows, integer overflows, UAF) are especially well-suited to FuzzTest, and which are harder to reach with this style of in-process fuzzing?

4. How could you integrate short, time-bounded FuzzTest runs into a CI pipeline without making builds too slow, while still having longer campaigns on dedicated fuzzing machines?

5. When writing properties like LengthFieldRespected, what kinds of mistakes in the property itself might cause you to miss real bugs or report lots of false positives?

Day 4: Introduction to Honggfuzz

• Goal: Understand different fuzzing methods and when to use Honggfuzz vs AFL++.

• Activities:

  • Reading: Continue with "Fuzzing for Software Security Testing and Quality Assurance" (From 5.1.2 to 5.3.7).

  • Online Resource: Honggfuzz

  • Real-World Context:

    • Honggfuzz is used in Google's continuous fuzzing infrastructure

    • OSS-Fuzz uses Honggfuzz for many projects

    • OpenSSL has extensive fuzzing coverage - see their fuzzing documentation

  • Exercise: Fuzz OpenSSL server and private key parsing

# Install Honggfuzz
cd ~/soft && git clone --branch master --depth 1 https://github.com/google/honggfuzz.git
#sudo apt-get install -y binutils-dev libunwind-dev libblocksruntime-dev clang libssl-dev
sudo apt-get install -y binutils-dev libunwind-dev libblocksruntime-dev libssl-dev
cd honggfuzz && make -j$(nproc) && sudo make install

# Verify installation
which honggfuzz
honggfuzz --version

# Clone OpenSSL source
cd ~/tuts && git clone --branch openssl-3.1.2 --depth=1 https://github.com/openssl/openssl.git
cd openssl

# Configure OpenSSL for fuzzing (with all legacy protocols enabled for broader attack surface)
CC=/usr/local/bin/hfuzz-clang CXX="$CC"++ ./config \
  -DPEDANTIC no-shared -DFUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION -O0 \
  -fno-sanitize=alignment -lm -ggdb -gdwarf-4 --debug -fno-omit-frame-pointer \
  enable-asan enable-tls1_3 enable-weak-ssl-ciphers enable-rc5 enable-md2 \
  enable-ssl3 enable-ssl3-method enable-nextprotoneg enable-heartbeats \
  enable-aria enable-zlib enable-egd

# Build OpenSSL
make -j$(nproc)

# Build Honggfuzz fuzzers for OpenSSL
# Note: This uses Honggfuzz's example OpenSSL fuzzers
cat > build_fuzzers.sh << 'EOT'
#!/bin/bash
set -x
set -e
echo "Building honggfuzz fuzzers for OpenSSL"
for x in x509 privkey client server; do
    hfuzz-clang \
        -DBORINGSSL_UNSAFE_DETERMINISTIC_MODE \
        -DBORINGSSL_UNSAFE_FUZZER_MODE \
        -DFUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION \
        -DBN_DEBUG \
        -DLIBRESSL_HAS_TLS1_3 \
        -O3 -g \
        -DFuzzerInitialize=LLVMFuzzerInitialize \
        -DFuzzerTestOneInput=LLVMFuzzerTestOneInput \
        -I$(pwd)/include \
        -I"$HOME"/soft/honggfuzz/examples/openssl \
        -I"$HOME"/soft/honggfuzz \
        -g "$HOME/soft/honggfuzz/examples/openssl/$x.c" \
        -o "libfuzzer.openssl-memory.$x" \
        ./libssl.a ./libcrypto.a -lpthread -lz -ldl -fsanitize=address
done
EOT

chmod +x build_fuzzers.sh
./build_fuzzers.sh

# Run Honggfuzz on OpenSSL server parser
honggfuzz --input ~/soft/honggfuzz/examples/openssl/corpus_server/ \
          -- ./libfuzzer.openssl-memory.server

# Run Honggfuzz on OpenSSL private key parser
honggfuzz --input ~/soft/honggfuzz/examples/openssl/corpus_privkey/ \
          -- ./libfuzzer.openssl-memory.privkey

Success Criteria:

• Honggfuzz compiles and installs successfully

• OpenSSL builds with fuzzing support

• Fuzzers compile without errors

• Honggfuzz starts and shows coverage statistics

What to Look For:

• Coverage metrics increasing over time

• Crashes in the working directory

• Different crash types (heap overflow, use-after-free, etc.)

Note: Real OpenSSL fuzzing often runs for days/weeks. For this exercise, run for at least 30 minutes to see initial results.

Real-World Impact: Honggfuzz Finding TLS Vulnerabilities

Case Study - Heartbleed-Class Bugs in TLS Implementations:

While Heartbleed (CVE-2014-0160) predates modern fuzzing tools, similar vulnerabilities continue to be found through continuous fuzzing campaigns.

Why TLS is Hard to Fuzz:

• Stateful protocol: Must complete handshake before reaching deep logic

• Cryptographic operations: Random values, signatures, MACs

• Multiple versions: TLS 1.0, 1.1, 1.2, 1.3 with different code paths

• Extensions: ALPN, SNI, session tickets, early data, etc.

Honggfuzz Advantages for Network Protocols:

# Example: Fuzzing TLS 1.3 handshake
cat > fuzz_tls13_handshake.c << 'EOF'
#include <openssl/ssl.h>
#include <openssl/err.h>
#include <stdint.h>
#include <stddef.h>

int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
    SSL_CTX *ctx = SSL_CTX_new(TLS_server_method());
    if (!ctx) return 0;

    SSL_CTX_set_min_proto_version(ctx, TLS1_3_VERSION);
    SSL_CTX_set_max_proto_version(ctx, TLS1_3_VERSION);

    BIO *in_bio = BIO_new(BIO_s_mem());
    BIO *out_bio = BIO_new(BIO_s_mem());

    SSL *ssl = SSL_new(ctx);
    SSL_set_bio(ssl, in_bio, out_bio);
    SSL_set_accept_state(ssl);

    BIO_write(in_bio, data, size);
    SSL_do_handshake(ssl);

    SSL_free(ssl);
    SSL_CTX_free(ctx);
    return 0;
}
EOF

# Compile with Honggfuzz
hfuzz-clang fuzz_tls13_handshake.c \
    -I"$HOME"/tuts/openssl/include \
    -L"$HOME"/tuts/openssl/lib \
    -lssl -lcrypto \
    -fsanitize=address,undefined \
    -o fuzz_tls13

# Run with Honggfuzz
# TODO: use a better corpus to actually find the vulnerabilities in that code
honggfuzz --input ~/soft/honggfuzz/examples/openssl/corpus_server/ \
          --threads 8 \
          --timeout 5 \
          -- ./fuzz_tls13

Real Bugs Found by Protocol Fuzzing:

From OpenSSL and other TLS implementations:

• Buffer overflows in certificate parsing: X.509 extension handling

• Use-after-free in session resumption: Ticket lifetime management

• Integer overflows in record layer: Length calculations

• State confusion bugs: Unexpected message ordering

Example: CVE-2022-0778 (OpenSSL Infinite Loop):

# Bug: Infinite loop in BN_mod_sqrt() when parsing elliptic curve points
# Discovery: Fuzzing certificate parsing with malformed EC parameters
# Impact: DoS via crafted certificate
# Fixed: OpenSSL 3.0.2, 1.1.1n

# Fuzzing campaign that found similar bugs:
honggfuzz --input certs/ \
          --dict openssl.dict \
          --threads 16 \
          --timeout 10 \
          --rlimit_rss 2048 \
          -- ./openssl x509 -in ___FILE___ -text

# Result: Timeout on malformed EC point → DoS vulnerability

Fuzzing vs Real-World Exposure:

Metric	Production Use (10 years)	OSS-Fuzz (1 year)
Total connections	Billions	0 (pure fuzzing)
Unique inputs tested	~1,000 (typical sites)	Trillions
Edge cases covered	<1%	>90%
Bugs found	~5 (via exploits)	~50

Key Insight: Fuzzing explores input space breadth that production traffic never reaches.

Key Takeaways

1. Honggfuzz excels at complex targets: Multi-threaded, persistent mode, hardware-assisted coverage

2. Protocol fuzzing requires stateful harnesses: Must reach deep code paths beyond initial parsing

3. Continuous fuzzing prevents regressions: OSS-Fuzz runs 24/7, catches new bugs in code changes

4. Cryptographic code is fragile: Parsers for ASN.1, X.509, PEM frequently have bugs

5. Timeout detection finds DoS bugs: Infinite loops, algorithmic complexity issues

Discussion Questions

1. Why does fuzzing find TLS bugs that years of production use don't reveal?

2. What makes protocol fuzzing (TLS, HTTP/2, DNS) more challenging than file format fuzzing?

3. How does hardware-assisted coverage (Intel PT) improve fuzzing effectiveness?

4. What are the limitations of fuzzing for finding cryptographic vulnerabilities vs implementation bugs?

Day 5: Introduction to Syzkaller

• Goal: Begin kernel fuzzing with Syzkaller.

• Activities:

  • Tool: Install Syzkaller on a Linux VM.

  • Online Resource: Syzkaller Documentation

  • Real-World Impact:

    • Syzkaller Dashboard - Shows thousands of bugs found by syzkaller

    • syzkaller: Finding Bugs in the Linux Kernel - Overview of syzkaller's impact

    • Many CVEs discovered: Check syzkaller bug reports for real examples

  • Exercise: Start fuzzing the Linux kernel with Syzkaller.

# Install kernel build dependencies
sudo apt update
sudo apt install -y make gcc flex bison libncurses-dev libelf-dev libssl-dev

# Clone Linux kernel (use a recent stable version)
# Check available tags: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/refs/tags
cd ~/soft && git clone --branch v6.8 --depth 1 git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git kernel

# Verify kernel version
cd kernel && git describe --tags

# Configure kernel for syzkaller
make defconfig
make kvm_guest.config

# Edit .config to enable syzkaller requirements
# Use sed or manually edit .config
sed -i 's/# CONFIG_KCOV is not set/CONFIG_KCOV=y/' .config
sed -i 's/# CONFIG_DEBUG_INFO_DWARF4 is not set/CONFIG_DEBUG_INFO_DWARF4=y/' .config
sed -i 's/# CONFIG_KASAN is not set/CONFIG_KASAN=y/' .config
sed -i 's/# CONFIG_KASAN_INLINE is not set/CONFIG_KASAN_INLINE=y/' .config
sed -i 's/# CONFIG_CONFIGFS_FS is not set/CONFIG_CONFIGFS_FS=y/' .config
sed -i 's/# CONFIG_SECURITYFS is not set/CONFIG_SECURITYFS=y/' .config
echo 'CONFIG_CMDLINE_BOOL=y' >> .config
echo 'CONFIG_CMDLINE="net.ifnames=0"' >> .config

make olddefconfig
make -j$(nproc)

# Create VM image
sudo apt install -y debootstrap
mkdir -p ~/soft/image && cd ~/soft/image
wget https://raw.githubusercontent.com/google/syzkaller/master/tools/create-image.sh -O create-image.sh
chmod +x create-image.sh
./create-image.sh --distribution trixie --feature full

# Install QEMU
sudo apt install -y qemu-system-x86

# Test VM boot (optional - verify image works)
# NOTE: You might need to run this outside of the vm, you might need to change net to e1000
cd /tmp/
sudo qemu-system-x86_64 \
    -m 2G -smp 2 \
    -kernel ~/soft/kernel/arch/x86/boot/bzImage \
    -append "console=ttyS0 root=/dev/sda earlyprintk=serial net.ifnames=0" \
    -drive file=~/soft/image/trixie.img,format=raw \
    -net user,hostfwd=tcp:127.0.0.1:10021-:22 \
    -net nic,model=virtio -enable-kvm -nographic \
    -pidfile vm.pid 2>&1 | tee vm.log

# In another terminal, test SSH access
ssh -i ~/soft/image/trixie.id_rsa -p 10021 -o "StrictHostKeyChecking no" root@localhost

# Install Go
cd ~/soft
GO_VERSION="1.25.4"
wget "https://go.dev/dl/go${GO_VERSION}.linux-amd64.tar.gz"
sudo tar -C /usr/local -xzf "go${GO_VERSION}.linux-amd64.tar.gz"
export PATH=$PATH:/usr/local/go/bin

# Add to ~/.bashrc for persistence
echo 'export PATH=$PATH:/usr/local/go/bin' >> ~/.bashrc

# Verify Go installation
go version

# Clone and build syzkaller
cd ~/soft && git clone --branch master --depth 1 https://github.com/google/syzkaller.git
cd syzkaller && make -j$(nproc)

# Create syzkaller configuration
# NOTE: If you are in the vm with e1000 then you don't need network_device line
cat > my.cfg << 'EOT'
{
    "target": "linux/amd64",
    "http": "127.0.0.1:56741",
    "workdir": "/home/USER/soft/syzkaller/workdir",
    "kernel_src": "/home/USER/soft/kernel",
    "image": "/home/USER/soft/image/trixie.img",
    "sshkey": "/home/USER/soft/image/trixie.id_rsa",
    "ssh_user": "root",
    "syzkaller": "/home/USER/soft/syzkaller",
    "procs": 8,
    "type": "qemu",
    "vm": {
        "count": 3,
        "kernel": "/home/USER/soft/kernel/arch/x86/boot/bzImage",
        "cmdline": "net.ifnames=0",
        "cpu": 2,
        "mem": 2048,
        "network_device": "virtio-net-pci"
    }
}
EOT

# Replace $USER with actual username
sed -i "s/USER/$USER/g" my.cfg

# Create workdir and start syzkaller
mkdir -p workdir
# NOTE: this might take a while, run with -debug to to identify issues
sudo ./bin/syz-manager -config=my.cfg

# Access web interface
# Install text-based browser or use your regular browser
sudo apt install -y w3m w3m-img
w3m http://127.0.0.1:56741

# Or open in regular browser: http://127.0.0.1:56741

Success Criteria:

• Kernel compiles successfully with KASAN and KCOV enabled

• VM image creates without errors

• VM boots and is accessible via SSH

• Syzkaller manager starts and shows web interface

• Web interface displays fuzzing statistics

Expected Outputs:

• Web interface showing: exec total, crashes, coverage, etc.

• Crashes appearing in workdir/crashes/ directory

• Kernel oops messages in VM logs

Troubleshooting:

• If kernel doesn't boot: Check QEMU/KVM is enabled (lsmod | grep kvm)

• If syzkaller can't connect: Verify SSH key permissions (chmod 600 trixie.id_rsa)

• If no crashes: Let run longer - kernel fuzzing takes time

• Memory issues: Reduce VM count in config if system runs out of RAM

Real-World Impact: Syzkaller's Contribution to Kernel Security

Case Study - CVE-2022-32250 (Linux Netfilter Use-After-Free):

From Week 1, you learned about this vulnerability. Here's how syzkaller discovered it:

• Target: net/netfilter/nf_tables_api.c - Linux firewall subsystem

• Discovery Date: May 2022

• Fuzzing Duration: ~72 hours from code introduction to crash

• Root Cause: Reference counting error in stateful expression handling

The Discovery Process:

# Syzkaller's approach (simplified)
# 1. System call description for netfilter operations
{
    "nfnetlink_create": {
        "protocol": "NETLINK_NETFILTER",
        "operations": ["NFT_MSG_NEWTABLE", "NFT_MSG_NEWCHAIN", "NFT_MSG_NEWRULE"]
    }
}

# 2. Syscall sequence that triggered the bug
socket(AF_NETLINK, SOCK_RAW, NETLINK_NETFILTER)
sendmsg(fd, {
    type: NFT_MSG_NEWTABLE,
    flags: NLM_F_CREATE | NLM_F_EXCL,
    data: [table_attrs]
})
sendmsg(fd, {
    type: NFT_MSG_NEWCHAIN,
    data: [chain_with_stateful_expr]
})
# Trigger: Modify stateful expression in specific sequence
sendmsg(fd, {
    type: NFT_MSG_NEWRULE,
    data: [rule_update_that_frees_expr]
})
# Use freed expression -> UAF crash

# 3. KASAN detected use-after-free
# BUG: KASAN: use-after-free in nf_tables_expr_destroy+0x12/0x20
# Read of size 8 at addr ffff888012345678 by task syz-executor/1234

Why Syzkaller Found It:

1. Syscall coverage: Tests all netfilter operations systematically

2. Sequence exploration: Tries millions of syscall orderings

3. State tracking: Maintains kernel state across operations

4. KASAN integration: Immediate detection of memory corruption

5. Reproducibility: Generates C reproducer for developers

The Reproducer (simplified):

// Generated by syzkaller - minimal reproducer
#include <sys/socket.h>
#include <linux/netlink.h>
#include <linux/netfilter/nf_tables.h>

int main(void) {
    int fd = socket(AF_NETLINK, SOCK_RAW, NETLINK_NETFILTER);

    // Create table
    send_nft_msg(fd, NFT_MSG_NEWTABLE, ...);

    // Create chain with stateful expression
    send_nft_msg(fd, NFT_MSG_NEWCHAIN, ...);

    // Trigger UAF through rule update
    send_nft_msg(fd, NFT_MSG_NEWRULE, ...);

    return 0;
}

Impact: Local privilege escalation from any user to root on systems with unprivileged user namespaces (default on Ubuntu, Debian). Public exploit available within weeks.

Case Study - CVE-2023-32629 (Linux Netfilter Race Condition):

• Target: net/netfilter/nf_tables_api.c - Same subsystem, different bug

• Bug Class: Race condition in batch transaction handling

• Discovery: Syzkaller's multi-threaded syscall fuzzing

• Impact: Container escape + privilege escalation

How Syzkaller Finds Race Conditions:

# Syzkaller executes syscalls in parallel across multiple threads
# VM 1, Thread 1:
socket(AF_NETLINK, ...) → fd1
sendmsg(fd1, NFT_MSG_NEWTABLE, ...)

# VM 1, Thread 2 (simultaneous):
socket(AF_NETLINK, ...) → fd2
sendmsg(fd2, NFT_MSG_NEWTABLE, ...)  # Race on same table

# Result: Concurrent access to nf_tables objects without proper locking
# KASAN detects: use-after-free or memory corruption

Syzkaller's Advantages for Kernel Fuzzing:

1. Syscall descriptions: Domain-specific language for kernel APIs

2. Coverage-guided: Tracks code coverage to explore new paths

3. Multi-threaded: Finds race conditions naturally

4. VM-based isolation: Kernel crashes don't affect fuzzer

5. Reproducers: Automatic generation of minimal C reproducers

6. Bisection: Automatically finds introducing commit

Analyzing a Syzkaller Bug:

# Download reproducer from dashboard
mkdir -p ~/tuts/syz-repro && cd ~/tuts/syz-repro
wget "https://syzkaller.appspot.com/text?tag=ReproC&x=15ad9542580000" -O repro.c

# Compile and test on vulnerable kernel
gcc -pthread -o repro repro.c
./repro

# Expected: Segmentation Fault
# Kernel log shows UBSAN: array-index-out-of-bounds

# Analyze how it got identified:
https://syzkaller.appspot.com/bug?extid=77026564530dbc29b854

Key Insight: Kernel attack surface is massive. Syzkaller's systematic approach finds bugs that would take years of manual testing.

Key Takeaways

1. Syzkaller revolutionized kernel security: Found 4,500+ bugs that manual testing missed

2. Syscall fuzzing requires domain knowledge: Must understand kernel APIs to fuzz effectively

3. Race conditions need parallel execution: Multi-threaded fuzzing essential

4. VM isolation is critical: Kernel crashes would kill the fuzzer otherwise

5. Reproducers enable fixing: Minimal C programs allow developers to debug quickly

Discussion Questions

1. Why has syzkaller found thousands of kernel bugs that years of production use didn't reveal?

2. How does syzkaller's syscall description language enable effective kernel fuzzing?

3. What makes race condition detection particularly valuable in kernel fuzzing?

4. Why are networking subsystems (netfilter, inet) the most frequent sources of vulnerabilities?

5. How do user namespaces make kernel vulnerabilities more dangerous by increasing exploitability?

Day 6: Crash Analysis and Exploitability Assessment

• Goal: Understand crash triage, root cause analysis, and exploitability assessment for fuzzer-discovered bugs.

• Activities:

  • Reading:

    • "Fuzzing for Software Security Testing and Quality Assurance" by Ari Takanen (Sections 6.1 to 6.6.6)

    • "The Art of Software Security Assessment" Chapter 5 and 6

  • Online Resources:

    • AddressSanitizer Documentation

    • CASR - Crash Analysis and Severity Report

    • OSS-Fuzz Guide

  • Real-World Context:

    • Exploitability Ratings

    • Crash Accumulation During Fuzzing with CASR

    • Effective Fuzzing Harness

  • Concepts:

    • Crash deduplication and bucketing

    • Root cause analysis from stack traces

    • Exploitability classification

    • ASAN report interpretation

    • Building proof-of-concept exploits

Understanding Crash Analysis Tools

# Install crash analysis toolkit
sudo apt update
sudo apt install -y gdb python3-pip valgrind binutils

# Install GEF (GDB Enhanced Features) - if you haven't done so already
#wget -O ~/.gdbinit-gef.py -q https://gef.blah.cat/py
#echo "source ~/.gdbinit-gef.py" >> ~/.gdbinit

# Install CASR (Crash Analysis and Severity Report) and rust if you haven't already
#curl --proto '=https' --tlsv1.2 -sSf "https://sh.rustup.rs" | sh
#source ~/.cargo/env
cargo install casr

# Verify installations
casr-san --help | head -5

Case Study 1: Analyzing Heap Buffer Overflow

Scenario: Fuzzing discovered a crash in an image parser. Let's perform complete analysis.

# Create sample vulnerable image parser
mkdir -p ~/crash_analysis/case1_heap_overflow && cd ~/crash_analysis/case1_heap_overflow

cat > vuln_parser.c << 'EOF'
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>

void build_huffman_table(uint8_t *input, size_t size) {
    if (size < 8) return;

    uint32_t table_size = *(uint32_t*)input;
    uint8_t *codes = input + 4;

    uint8_t *table = malloc(256);

    // VULNERABILITY: No bounds check on table_size
    // Can write beyond 256-byte buffer
    memcpy(table, codes, table_size);  // Heap buffer overflow!

    printf("Built Huffman table with %u codes\n", table_size);

    free(table);
}

int main(int argc, char **argv) {
    if (argc < 2) return 1;

    FILE *f = fopen(argv[1], "rb");
    if (!f) return 1;

    fseek(f, 0, SEEK_END);
    size_t size = ftell(f);
    fseek(f, 0, SEEK_SET);

    uint8_t *data = malloc(size);
    fread(data, 1, size, f);
    fclose(f);

    build_huffman_table(data, size);

    free(data);
    return 0;
}
EOF

# Compile with ASAN for detailed crash reports
clang-19 -g -O0 -fsanitize=address -o vuln_parser_asan vuln_parser.c

# Create crashing input (table_size = 512, overflows 256-byte buffer)
python3 << 'EOF'
import struct
# table_size = 512 (causes 256-byte overflow)
payload = struct.pack('<I', 512)
payload += b'A' * 512
with open('crash_heap_overflow.bin', 'wb') as f:
    f.write(payload)
EOF

# Run and capture ASAN output
./vuln_parser_asan crash_heap_overflow.bin 2>&1 | tee asan_crash.log

ASAN Output Analysis:

==37160==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x511000000140 at pc 0x56d6a37d0f62 bp 0x7ffd9f024440 sp 0x7ffd9f023c00
WRITE of size 512 at 0x511000000140 thread T0
    #0 0x56d6a37d0f61 in __asan_memcpy
    #1 0x56d6a38147f5 in build_huffman_table vuln_parser.c:16:5
    #2 0x56d6a38148fe in main vuln_parser.c:37:5

0x511000000140 is located 0 bytes after 256-byte region [0x511000000040,0x511000000140)
allocated by thread T0 here:
    #0 0x56d6a37d3193 in malloc (vuln_parser_asan+0xcc193) (BuildId: e524ec295f274ddf6e407b3941080060bdfc9d1c)
    #1 0x56d6a38147df in build_huffman_table vuln_parser.c:12:22
    #2 0x56d6a38148fe in main vuln_parser.c:37:5

SUMMARY: AddressSanitizer: heap-buffer-overflow (vuln_parser_asan+0xc9f61) (BuildId: e524ec295f274ddf6e407b3941080060bdfc9d1c) in __asan_memcpy
Shadow bytes around the buggy address:
  0x510ffffffe80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x510fffffff00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x510fffffff80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x511000000000: fa fa fa fa fa fa fa fa 00 00 00 00 00 00 00 00
  0x511000000080: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
=>0x511000000100: 00 00 00 00 00 00 00 00[fa]fa fa fa fa fa fa fa

Interpreting the ASAN Report:

1. Bug Type: heap-buffer-overflow

2. Operation: WRITE of size 512

3. Location: vuln_parser.c:16 in build_huffman_table()

4. Allocation: 256-byte buffer at line 12

5. Overflow: Writing 512 bytes into 256-byte buffer = 256 bytes overflow

Root Cause Analysis:

# View the vulnerable code with context
cat -n vuln_parser.c | sed -n '6,16p'
#     6  void build_huffman_table(uint8_t *input, size_t size) {
#     7      if (size < 8) return;
#     8
#     9      uint32_t table_size = *(uint32_t*)input;  // ATTACKER CONTROLLED
#    10      uint8_t *codes = input + 4;
#    11
#    12      uint8_t *table = malloc(256);             // Fixed 256 bytes
#    13
#    14      // VULNERABILITY: No bounds check on table_size
#    15      // Can write beyond 256-byte buffer
#    16      memcpy(table, codes, table_size);         // Copies attacker-controlled amount!

Exploitability Assessment:

# Classify crash automatically with CASR (using the ASAN run you captured above)
#casr-san -o heap_overflow.casrep -- ./vuln_parser_asan crash_heap_overflow.bin 2>&1 | tee casr_heap_overflow.log
casr-san --stdout -- ./vuln_parser_asan crash_heap_overflow.bin | tee heap_overflow.casrep

# The .casrep and log will contain fields like:
#   "Type": "EXPLOITABLE",
#   "ShortDescription": "heap-buffer-overflow(write)",

# You can still use GDB for manual inspection of the corrupted heap if you want:
gdb ./vuln_parser_asan
(gdb) run crash_heap_overflow.bin

Exploitability Classification: EXPLOITABLE

Reasoning:

1. Attacker controls overflow size: table_size from input

2. Attacker controls overflow data: codes array content

3. Heap corruption possible: Can overwrite adjacent objects

4. Exploitation path:

   • Overflow into adjacent heap object

   • Corrupt function pointer or vtable

   • Hijack control flow

   • Execute arbitrary code

Real-World Example: Similar to CVE-2023-4863 (libWebP Heap Buffer Overflow) from Week 1.

Case Study 2: Use-After-Free Analysis

mkdir -p ~/crash_analysis/case2_uaf && cd ~/crash_analysis/case2_uaf

cat > vuln_uaf.c << 'EOF'
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

typedef struct {
    char *name;
    void (*process)(void);
} Handler;

void default_handler(void) {
    printf("Default handler\n");
}

void evil_handler(void) {
    printf("Evil handler executed! Code execution via UAF.\n");
}

Handler *handler = NULL;

void register_handler(char *name) {
    handler = malloc(sizeof(Handler));
    handler->name = strdup(name);
    handler->process = default_handler;
}

void unregister_handler(void) {
    if (handler) {
        free(handler->name);
        free(handler);
        // BUG: Should set handler = NULL here!
    }
}

void attacker_groom_heap(void) {
    for (int i = 0; i < 1000; i++) {
        Handler *fake = malloc(sizeof(Handler));
        fake->name = "pwned";
        fake->process = evil_handler;
    }
}

void call_handler(void) {
    if (handler) {
        handler->process();
    }
}

int main(int argc, char **argv) {
    register_handler("test");
    unregister_handler();
    attacker_groom_heap();
    call_handler();

    return 0;
}
EOF

# Compile with ASAN
clang-19 -g -O0 -fsanitize=address -o vuln_uaf_asan vuln_uaf.c

# Run and capture output
./vuln_uaf_asan 2>&1 | tee uaf_crash.log

ASAN Output:

=================================================================
==38664==ERROR: AddressSanitizer: heap-use-after-free on address 0x502000000010 at pc 0x617b2245a953 bp 0x7ffe92f7c160 sp 0x7ffe92f7c158
READ of size 8 at 0x502000000010 thread T0
    #0 0x617b2245a952 in call_handler vuln_uaf.c:44:50
    #1 0x617b2245a9d0 in main vuln_uaf.c:53:5

0x502000000010 is located 0 bytes inside of 16-byte region [0x502000000010,0x502000000020)
freed by thread T0 here:
    #1 0x617b2245a86a in unregister_handler vuln_uaf.c:29:9
    #2 0x617b2245a9c6 in main vuln_uaf.c:51:5

previously allocated by thread T0 here:
    #1 0x617b2245a7a5 in register_handler vuln_uaf.c:21:15
    #2 0x617b2245a9c1 in main vuln_uaf.c:50:5

SUMMARY: AddressSanitizer: heap-use-after-free vuln_uaf.c:44:50 in call_handler
Shadow bytes around the buggy address:
  0x501ffffffd80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x501ffffffe00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x501ffffffe80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x501fffffff00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x501fffffff80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
=>0x502000000000: fa fa[fd]fd fa fa fd fa fa fa 00 00 fa fa 00 00

Exploitability Assessment:

# Classify crash with CASR again (now for a UAF instead of overflow)
casr-san --stdout -- ./vuln_uaf_asan 2>&1 | tee uaf.casrep

# The report will highlight:
#   - Bug class: heap-use-after-free
#   - Severity: "NOT_EXPLOITABLE"
#   - Reason: ASan instruments the READ of the function pointer *before* the call.
#     Since it's a read from freed memory, CASR defaults to "Not Exploitable".
#
#     This is a critical lesson: Automated tools are heuristics.
#     A human analyst sees "Read of function pointer from freed memory" -> Critical.

# To prove exploitability, we would need to:
# 1. Bypass ASan quarantine (so memory is reallocated/valid)
# 2. Overwrite with a bad pointer
# 3. Trigger the crash on the JUMP (SEGV), not the UAF read.
#
# For now, trust your manual analysis: Controlling a function pointer is exploitable.

# For deeper debugging, verify the exploit manually in GDB:
# NOTE: You must disable ASan quarantine to allow the freed chunk to be reused!
# Otherwise, malloc() will return a new address, and handler->process will still be old/garbage.
export ASAN_OPTIONS=detect_leaks=0:quarantine_size_mb=0

gdb ./vuln_uaf_asan
(gdb) break vuln_uaf.c:44   # Break at call_handler (before crash)
(gdb) run
(gdb) p handler
# $1 = (Handler *) 0x...

(gdb) p *handler
# With quarantine=0, you should see:
# name = "pwned"
# process = <evil_handler>

(gdb) p handler->process
# Should point to evil_handler

(gdb) continue
# Execution should flow to evil_handler() (or crash if ASan still catches the shadow marker)

Exploitability Classification: EXPLOITABLE (Verified via manual analysis)

[!NOTE]: Automated tools like CASR may label this NOT_EXPLOITABLE because ASan instruments the function pointer read before the call. Manual verification (as shown above) proves control flow hijack is possible.

Exploitation Strategy:

1. Heap grooming: Allocate/free to position objects

2. Reclaim freed memory: Allocate object of same size (requires bypassing ASan quarantine in lab)

3. Control freed memory contents: Fill with attacker data

4. Trigger UAF: Call call_handler()

5. Function pointer hijack: handler->process points to attacker-controlled address

6. Result: Arbitrary code execution

Real-World Example: Similar to CVE-2024-2883 (Chrome ANGLE UAF) from Week 1.

Case Study 3: Integer Overflow Leading to Heap Corruption

mkdir -p ~/crash_analysis/case3_integer_overflow && cd ~/crash_analysis/case3_integer_overflow

cat > vuln_intoverflow.c << 'EOF'
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <stdint.h>

void process_image(uint32_t width, uint32_t height, uint8_t *data) {
    size_t pixel_count = width * height;
    size_t buffer_size = pixel_count * 4;

    printf("Allocating %zu bytes for %ux%u image\n", buffer_size, width, height);

    uint8_t *buffer = malloc(buffer_size);

    for (size_t i = 0; i < (size_t)width * height; i++) {
        // WRITES out of bounds immediately
        // Use modulo to avoid reading out of bounds of 'data'
        buffer[i * 4] = data[i % 1024];
    }

    free(buffer);
}

int main(int argc, char **argv) {
    // Attacker-controlled dimensions
    uint32_t width = 0x10000;   // 65536
    uint32_t height = 0x10000;  // 65536

    uint8_t fake_data[1024];
    memset(fake_data, 'A', sizeof(fake_data));

    process_image(width, height, fake_data);

    return 0;
}
EOF

# Compile with ASAN and UBSAN
clang-19 -g -O0 -fsanitize=address,unsigned-integer-overflow -o vuln_int_asan vuln_intoverflow.c

# Run
./vuln_int_asan 2>&1 | tee intoverflow_crash.log

Analysis:

vuln_intoverflow.c:7:32: runtime error: unsigned integer overflow: 65536 * 65536 cannot be represented in type 'uint32_t' (aka 'unsigned int')
SUMMARY: UndefinedBehaviorSanitizer: undefined-behavior vuln_intoverflow.c:7:32
=================================================================
==39011==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x502000000014 at pc 0x5fa5104bd933 bp 0x7ffd7d3885a0 sp 0x7ffd7d388598
WRITE of size 1 at 0x502000000014 thread T0
    #0 0x5fa5104bd932 in process_image vuln_intoverflow.c:17:23
    #1 0x5fa5104bdad0 in main vuln_intoverflow.c:31:5

0x502000000014 is located 3 bytes after 1-byte region [0x502000000010,0x502000000011)
allocated by thread T0 here:
    #1 0x5fa5104bd7fc in process_image vuln_intoverflow.c:12:23

SUMMARY: AddressSanitizer: heap-buffer-overflow vuln_intoverflow.c:17:23 in process_image
Shadow bytes around the buggy address:
  0x501ffffffd80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x501ffffffe00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x501ffffffe80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x501fffffff00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x501fffffff80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
=>0x502000000000: fa fa[01]fa fa fa fa fa fa fa fa fa fa fa fa fa

Root Cause:

1. Integer Overflow: width * height overflows 32-bit integer range, wrapping to 0.

2. Under-allocation: malloc(0) allocates a tiny chunk.

3. Logic Mismatch: Loop uses proper 64-bit bounds (or nested loops), iterating 4 billion times.

4. Heap Corruption: Loop writes far beyond the allocated chunk.

Exploitability Assessment:

# Classify crash with CASR
# Note: UBSAN might print an error first, but ASAN catches the memory corruption
casr-san --stdout -- ./vuln_int_asan 2>&1 | tee intoverflow.casrep

# The report will highlight:
#    "Type": "EXPLOITABLE",
#    "ShortDescription": "heap-buffer-overflow(write)",
#    "Description": "Heap buffer overflow",

# Verify manually in GDB:
gdb ./vuln_int_asan
(gdb) break vuln_intoverflow.c:17   # Break inside the loop
(gdb) run
(gdb) p buffer_size
# $1 = 0
(gdb) p buffer
# $2 = (uint8_t *) 0x... (Valid small chunk)
(gdb) x/4gx buffer
# Check adjacent memory (likely metadata or other chunks)
(gdb) step
# Watch the write to buffer[0] corrupting the heap
(gdb) continue

Exploitability Classification: EXPLOITABLE

Exploitation Strategy:

1. Heap Grooming: Allocate a sensitive object (e.g., a structure with a function pointer) immediately after the vulnerable 0-byte allocation.

2. Trigger Overflow: Send input with dimensions 0x10000 * 0x10000 to cause integer overflow -> malloc(0).

3. Overwrite: The loop writes attacker data (fake_data) into the adjacent sensitive object.

4. Hijack: Trigger the use of the corrupted object (e.g., call the function pointer).

Building Proof-of-Concept Exploits

Heap Overflow Example

cd ~/crash_analysis/case1_heap_overflow
cat > buf_ov.py << 'EOF'
#!/usr/bin/env python3
# POC for heap buffer overflow in vuln_parser

import struct

def craft_overflow_input():
    """Create input that causes a massive overflow"""
    # Target allocates 256 bytes. We claim size is 512.
    table_size = 512

    # Structure: [Size (4 bytes)] [Data (512 bytes)]
    payload = struct.pack('<I', table_size)
    payload += b'A' * table_size

    return payload

with open('poc_exploit.bin', 'wb') as f:
    f.write(craft_overflow_input())

print("[+] Created poc_exploit.bin")
print("[+] Run the ASan build for diagnostics or the no-sanitizer build for raw exploitation.")
EOF
python3 buf_ov.py
# Diagnostic run (shows detailed ASan report)
./vuln_parser_asan poc_exploit.bin

# Realistic exploit run (no sanitizers, glibc notices corrupted heap metadata)
clang-19 -g -O0 -o vuln_parser_nosan vuln_parser.c
MALLOC_CHECK_=0 ./vuln_parser_nosan poc_exploit.bin

Use-After-Free Example

cd ~/crash_analysis/case2_uaf

# Build a runtime version without sanitizers (ensures freed chunks are reused immediately)
clang-19 -g -O0 -o vuln_uaf_nosan vuln_uaf.c

# Optional: keep the ASan build for triage but disable quarantine to watch exploitation
# ASAN_OPTIONS=quarantine_size_mb=0 ./vuln_uaf_asan

# Execute the no-sanitizer build to watch the hijacked handler fire
./vuln_uaf_nosan

Expected console output:

Evil handler executed! Code execution via UAF.

To force an outright crash (showing instruction-pointer control), change the spray in attacker_groom_heap() to:

fake->process = (void (*)(void))0x4141414141414141ULL;

Running ./vuln_uaf_nosan now ends with a segfault at 0x4141414141414141, demonstrating control-flow hijack without AddressSanitizer interfering.

Key Takeaways

1. Sanitized vs non-sanitized builds: Use ASan/UBSan/KASAN builds for triage and root-cause, then switch to no-sanitizer builds (with knobs like ASAN_OPTIONS or MALLOC_CHECK_) to study realistic heap behavior and exploitation.

2. Automated ratings are heuristics: CASR’s Type/Severity fields are a starting point only; Case Study 2 showed a UAF rated NOT_EXPLOITABLE even though a function pointer hijack is clearly possible.

3. Crash location vs root cause: Tools often stop at the first invalid access (e.g., a read from freed memory) while the real exploit primitive (e.g., control-flow hijack) may be one instruction later.

4. Exploitability hinges on control: In all three case studies, exploitation becomes realistic when the attacker controls sizes (length, dimensions) and data that drive allocation and memory writes.

5. Systematic PoC development: The path is always fuzz → crash → triage (ASan + CASR) → root cause → minimal reproducer → exploit PoC (heap metadata or function pointer overwrite).

Discussion Questions

1. In your own workflow, when would you prefer to keep AddressSanitizer enabled, and when would you switch to a no-sanitizer build while evaluating exploitability?

2. How could CASR’s NOT_EXPLOITABLE rating for the UAF case mislead a less experienced analyst, and what manual checks (in GDB) prevent that mistake?

3. In the integer-overflow case, which variables and addresses would you inspect in GDB to confirm both under-allocation and the ensuing heap overwrite?

4. How does the exact crash site (e.g., first invalid read vs later jump through a corrupted function pointer) change your assessment of exploitability and which tools notice it?

5. How do modern mitigations (ASLR, DEP, hardened allocators, CFI) interact with the exploitation strategies you used in Case Studies 1–3, and what extra steps would be needed in a real target?

Further Reading

Blog Posts and Case Studies

• Leveling Up Fuzzing: Finding more vulnerabilities with AI

• AFL Success Stories

• 5 CVEs Found with Feedback-Based Fuzzing

• Syzkaller: Finding Bugs in the Linux Kernel

• OpenSSL Fuzzing Guide

• Slice: SAST + LLM Interprocedural Context Extractor - Using build-free CodeQL + Tree-Sitter + GPT‑5 to triage ~1700 static UAF candidates in the Linux kernel down to a single real bug (CVE-2025-37778), a good complement to fuzzing-based crash triage.

Practice Targets

• Fuzzing-Module - Learning exercises

• Damn Vulnerable C Program - Vulnerable code for practice

Day 7: Fuzzing Harness Development and Real-World Campaigns

• Goal: Learn to write effective fuzzing harnesses and understand real-world fuzzing campaigns.

• Activities:

  • Reading:

    • libFuzzer Tutorial

    • OSS-Fuzz Integration Guide

  • Online Resources:

    • Fuzzing Harness Examples

    • ClusterFuzz - Continuous fuzzing infrastructure

  • Concepts:

    • Harness design principles

    • In-process vs out-of-process fuzzing

    • Persistent mode optimization

    • Seed corpus curation

    • Continuous fuzzing integration

What is a Fuzzing Harness?

A fuzzing harness is the code that:

1. Receives fuzzer-generated input

2. Prepares that input for the target API

3. Calls the target functionality

4. Handles errors/cleanup

Example - Bad Harness vs Good Harness:

// BAD HARNESS: Slow, inefficient
int main(int argc, char **argv) {
    FILE *f = fopen(argv[1], "rb");  // File I/O every iteration!
    // ... read file ...
    // ... call target API ...
    fclose(f);
    return 0;
}

// GOOD HARNESS: Fast, in-process
extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
    // Direct memory buffer, no I/O
    // Called thousands of times per second in same process
    target_api(data, size);
    return 0;
}

Case Study: Writing Harness for JSON Parser

Target: json-c library (real-world JSON parser)

mkdir -p ~/harness_dev && cd ~/harness_dev && mkdir json_harness && cd json_harness

# Clone json-c
git clone --depth 1 https://github.com/json-c/json-c.git

cd ~/harness_dev/json_harness/json-c
export CC=clang-19
export CXX=clang++-19
export CFLAGS="-fno-sanitize-coverage=trace-cmp -fsanitize=fuzzer-no-link,address -g -O1"
export CXXFLAGS="-fno-sanitize-coverage=trace-cmp -fsanitize=fuzzer-no-link,address -g -O1"
cmake -DBUILD_SHARED_LIBS=OFF ./
make -j$(nproc)

cd ~/harness_dev/json_harness

# Create fuzzing harness
cat > fuzz_parse.c << 'EOF'
#include <json-c/json.h>
#include <stdint.h>
#include <stddef.h>

int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
    const char *data1 = (const char *)data;
    json_tokener *tok = json_tokener_new();
    json_object *obj = json_tokener_parse_ex(tok, data1, size);

    if (obj) {
        // Exercise different API functions to increase coverage
        json_object_to_json_string_ext(obj, JSON_C_TO_STRING_PRETTY | JSON_C_TO_STRING_SPACED);

        if (json_object_is_type(obj, json_type_object)) {
            json_object_object_foreach(obj, key, val) {
                (void)json_object_get_type(val);
                (void)json_object_get_string(val);
            }
        }

        if (json_object_is_type(obj, json_type_array)) {
            size_t len = json_object_array_length(obj);
            for (size_t i = 0; i < len; i++) {
                json_object_array_get_idx(obj, i);
            }
        }

        json_object_put(obj);
    }

    json_tokener_free(tok);
    return 0;
}
EOF

# Compile with libFuzzer
clang-19 -g -fsanitize=address,fuzzer \
    -fno-sanitize-coverage=trace-cmp \
    -I. -Ijson-c \
    fuzz_parse.c \
    json-c/libjson-c.a \
    -o fuzz_json

# Create seed corpus
mkdir -p corpus
cat > corpus/valid1.json << 'EOF'
{"name": "test", "value": 42}
EOF

cat > corpus/valid2.json << 'EOF'
[1, 2, 3, {"nested": "object"}]
EOF

cat > corpus/valid3.json << 'EOF'
{
    "string": "value",
    "number": 123,
    "boolean": true,
    "null": null,
    "array": [1, 2, 3],
    "object": {"key": "value"}
}
EOF

# Run libFuzzer, it might take a while to actually crash
./fuzz_json corpus/ -max_total_time=300 -print_final_stats=1 -max_len=10000

Harness Design Principles Applied

1. In-process execution: LLVMFuzzerTestOneInput - no fork/exec overhead

2. Direct API targeting: Calls json_tokener_parse_ex directly

3. Coverage maximization: Exercises multiple code paths (objects, arrays, serialization)

4. Proper cleanup: Frees allocated memory to avoid OOM

5. Sanitizer-friendly: Works with ASAN/UBSAN for bug detection

Case Study: Fuzzing Archive Extractors

While CVE-2023-38831 was in closed-source WinRAR, let's fuzz open-source alternatives with similar architectures.

cd ~/harness_dev && mkdir archive_campaign && cd archive_campaign

# Target: libarchive (used by many archive tools)
git clone --depth 1 --branch v3.5.1 https://github.com/libarchive/libarchive.git
cd ~/harness_dev/archive_campaign/libarchive

export CC=clang-19
export CXX=clang++-19
export CFLAGS="-fno-sanitize-coverage=trace-cmp -fsanitize=fuzzer-no-link,undefined,address -g -O1"
export CXXFLAGS="-fno-sanitize-coverage=trace-cmp -fsanitize=fuzzer-no-link,undefined,address -g -O1"

cmake -DENABLE_TEST=OFF .
make -j$(nproc)

cd ~/harness_dev/archive_campaign/

cat > fuzz_archive_read.c << 'EOF'
#include "archive.h"
#include "archive_entry.h"
#include <stdint.h>
#include <stddef.h>
#include <stdlib.h>

int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
    struct archive *a;
    struct archive_entry *entry;
    int r;

    a = archive_read_new();
    if (!a) return 0;

    archive_read_support_filter_all(a);
    archive_read_support_format_all(a);

    r = archive_read_open_memory(a, data, size);
    if (r != ARCHIVE_OK) {
        archive_read_free(a);
        return 0;
    }

    while (archive_read_next_header(a, &entry) == ARCHIVE_OK) {
        const char *name = archive_entry_pathname(entry);
        int64_t size = archive_entry_size(entry);
        mode_t mode = archive_entry_mode(entry);
        time_t mtime = archive_entry_mtime(entry);

        const char *linkname = archive_entry_symlink(entry);
        const char *hardlink = archive_entry_hardlink(entry);

        const void *buff;
        size_t read_size;
        int64_t offset;
        while (archive_read_data_block(a, &buff, &read_size, &offset) == ARCHIVE_OK) {
            // Just consume data, no processing needed
        }
    }

    archive_read_free(a);
    return 0;
}
EOF

# Compile harness
clang-19 -g -O1 -fsanitize=address,undefined,fuzzer \
    -fno-sanitize-coverage=trace-cmp \
    -I./libarchive/libarchive \
    fuzz_archive_read.c -o fuzz_archive \
    libarchive/libarchive/libarchive.a \
    -lz -llzma -lzstd -lxml2 -lcrypto -ldl -lpthread

# Build diverse seed corpus
mkdir -p corpus_archive

# Valid archives of different formats
# ZIP
echo "Test file" > /tmp/test.txt
zip corpus_archive/sample.zip /tmp/test.txt

# TAR.GZ
tar czf corpus_archive/sample.tar.gz /tmp/test.txt

# 7z (if available)
7z a corpus_archive/sample.7z /tmp/test.txt 2>/dev/null || true

# RAR (if available)
rar a corpus_archive/sample.rar /tmp/test.txt 2>/dev/null || true

# Archive with symlink (common bug location)
ln -sf /etc/passwd /tmp/symlink_test
tar czf corpus_archive/with_symlink.tar.gz /tmp/symlink_test 2>/dev/null || true

# Run fuzzing campaign
./fuzz_archive corpus_archive/ \
    -max_total_time=3600 \
    -timeout=30 \
    -rss_limit_mb=2048 \
    -print_final_stats=1

What This Campaign Targets:

1. Format parsing bugs: TAR, ZIP, RAR, 7z, etc.

2. Compression algorithms: gzip, bzip2, lzma, zstd

3. Path traversal: Symlink/hardlink handling (like CVE-2023-38831)

4. Metadata parsing: Timestamps, permissions, extended attributes

5. Memory corruption: Buffer overflows in decompression routines

Expected Findings (based on real OSS-Fuzz results):

• Integer overflows in size calculations

• Path traversal via symlinks

• Buffer overflows in compression codecs

• Use-after-free in error handling paths

Key Takeaways

1. Harness Design is Critical: Efficient harnesses (in-process, persistent) significantly outperform naive fork()/exec() wrappers.

2. Target Logic, Not Just I/O: Good harnesses bypass CLI parsing to exercise core API logic directly (e.g., json_tokener_parse_ex).

3. Seed Corpus Quality: A diverse, minimized corpus of valid inputs accelerates code coverage discovery.

4. Sanitizers Enable Detection: Memory bugs (ASAN) and undefined behavior (UBSAN) are only found if the harness is compiled with them.

5. Continuous Integration: Tools like OSS-Fuzz automate the "find → fix → verify" loop, preventing regressions in evolved code.

Discussion Questions

1. Why is an in-process harness (like LLVMFuzzerTestOneInput) orders of magnitude faster than a file-based CLI wrapper?

2. How does defining a proper seed corpus (e.g., valid JSON/ZIP files) help the fuzzer penetrate deeper into the target's logic?

3. What are the risks of "over-mocking" in a harness (e.g., bypassing too much initialization) versus "under-mocking" (doing too much I/O)?

4. How do you handle state cleanup in a persistent-mode harness to prevent false positives from memory leaks or global state pollution?

5. Why is it important to fuzz different layers of an application (e.g., the compression layer vs. the archive parsing layer) separately?

Week 2 Capstone Project: The Fuzzing Campaign

• Goal: Apply the week's techniques to discover and analyze a vulnerability in a real-world open source target or a complex challenge binary.

• Activities:

  • Select a Target:

    • Choose a C/C++ library that parses complex data (e.g., JSON, XML, Images, Archives, Network Packets).

    • Suggestions: json-c, libarchive, libpng, tinyxml2, mbedtls, or a known vulnerable version of a project (e.g., libwebp 1.0.0).

  • Harness Development:

    • Write a LLVMFuzzerTestOneInput harness or an AFL++ persistent mode harness.

    • Ensure the harness compiles with ASAN and UBSAN.

  • Campaign Execution:

    • Gather a valid seed corpus (from the internet or by generating samples).

    • Minimize the corpus using afl-cmin or afl-tmin (or libFuzzer's merge mode).

    • Run the fuzzer for at least 4 hours (or until a crash is found).

  • Triage and Analysis:

    • Deduplicate crashes.

    • Use GDB and ASAN reports to identify the root cause (e.g., Heap Overflow, UAF).

    • Determine exploitability (Control of instruction pointer? Arbitrary write?).

  • Report:

    • Document the target, harness code, campaign commands, and crash analysis.

Deliverables

• A fuzzing_report.md containing:

  • Target Details: Project name, version, and function targeted.

  • Harness Code: The C/C++ harness you wrote.

  • Campaign Stats: Fuzzer used, duration, executions/sec, and coverage achieved.

  • Crash Analysis: ASAN output, GDB investigation, and root cause explanation.

  • PoC: A minimal input file that triggers the crash.

Looking Ahead to Week 3

Next week, you'll learn patch diffing - analyzing security updates to understand what was fixed and discovering variant vulnerabilities. You'll see how fuzzing discoveries lead to patches, and how analyzing those patches can reveal additional bugs.