AFLGo: Directed Greybox Fuzzing

AFLGo is an extension of American Fuzzy Lop (AFL). Given a set of target locations (e.g., folder/file.c:582), AFLGo generates inputs specifically with the objective to exercise these target locations.

Unlike AFL, AFLGo spends most of its time budget on reaching specific target locations without wasting resources stressing unrelated program components. This is particularly interesting in the context of

patch testing by setting changed statements as targets. When a critical component is changed, we would like to check whether this introduced any vulnerabilities. AFLGo, a fuzzer that can focus on those changes, has a higher chance of exposing the regression.
static analysis report verification by setting statements as targets that a static analysis reports as potentially dangerous or vulnerability-inducing. When assessing the security of a program, static analysis tools might identify dangerous locations, such as critical system calls. AFLGo can generate inputs that actually show that this is indeed no false positive.
information flow detection by setting sensitive sources and sinks as targets. To expose data leakage vulnerabilities, a security researcher would like to generate executions that exercise sensitive sources containing private information and sensitive sinks where data becomes visible to the outside world. A directed fuzzer can be used to generate such executions efficiently.
crash reproduction by setting method calls in the stack-trace as targets. When in-field crashes are reported, only the stack-trace and some environmental parameters are sent to the in-house development team. To preserve the user's privacy, the specific crashing input is often not available. AFLGo could help the in-house team to swiftly reproduce these crashes.

AFLGo is based on AFL from Michał Zaleski <lcamtuf@coredump.cx>. Checkout the project awesome-directed-fuzzing for related work on directed greybox/whitebox fuzzing.

Getting Started

Let's get started with building AFLGo (on Ubuntu 20.04) and fuzz the target libxml2:

git clone https://github.com/aflgo/aflgo.git
cd aflgo
export AFLGO=$PWD

# Build AFLGo
sudo ./build.sh

# When you fuzz for the very first time...
sudo sh -c 'echo core > /proc/sys/kernel/core_pattern'

# Fuzz the target libxml2
cd examples
./libxml2-ef709ce2.sh

See the detailed steps discussed below.

Integration into OSS-Fuzz

The easiest way to use AFLGo is as patch testing tool in OSS-Fuzz. Here is our integration:

https://github.com/aflgo/oss-fuzz

Environment Variables

AFLGO_INST_RATIO -- The proportion of basic blocks instrumented with distance values (default: 100).
AFLGO_SELECTIVE -- Add AFL-trampoline only to basic blocks with distance values? (default: off).
AFLGO_PROFILER_FILE -- When CFG-tracing is enabled, the data will be stored here. (See instrument/README.md)

How to instrument a Binary with AFLGo

You can run AFLGo building script to do everything for you instead of manually go through step 1 to step 3. Be careful in these steps we would download, build and install LLVM 11.0.0 from source, which may have unexpected impacts on compiler toolchain in current system.

For step 4 to step 8, we are going to take libxml2 as an example. You can also equivalently run libxml2 fuzzing script instead.

Before we start, make sure that source code tree of AFLGo is ready and we are in its root. Then set the environment variable AFLGO to it, which will be used in later steps. For example,

git clone https://github.com/aflgo/aflgo.git
cd aflgo
export AFLGO=$PWD

Install LLVM 11.0.0 with Gold-plugin. Then make sure that the following commands successfully executed:

# Install LLVMgold into bfd-plugins
mkdir /usr/lib/bfd-plugins
cp /usr/local/lib/libLTO.so /usr/lib/bfd-plugins
cp /usr/local/lib/LLVMgold.so /usr/lib/bfd-plugins

Install other prerequisite

sudo apt-get update
sudo apt-get install python3
sudo apt-get install python3-dev
sudo apt-get install python3-pip
sudo apt-get install pkg-config
sudo apt-get install autoconf
sudo apt-get install automake
sudo apt-get install libtool-bin
sudo apt-get install gawk
sudo apt-get install libboost-all-dev  # boost is not required if you use gen_distance_orig.sh in step 7
python3 -m pip install networkx  # May vary by different python versions, see the case statement in build.sh
python3 -m pip install pydot
python3 -m pip install pydotplus

Compile AFLGo fuzzer, LLVM-instrumentation pass and the distance calculator

export CXX=`which clang++`
export CC=`which clang`
export LLVM_CONFIG=`which llvm-config`

pushd afl-2.57b; make clean all; popd;
pushd instrument; make clean all; popd;
pushd distance/distance_calculator; cmake ./; cmake --build ./; popd;

Download subject libxml2.

# Clone subject repository
git clone https://gitlab.gnome.org/GNOME/libxml2
export SUBJECT=$PWD/libxml2

Set targets (e.g., changed statements in commit ef709ce2). Writes BBtargets.txt.

# Setup directory containing all temporary files
mkdir temp
export TMP_DIR=$PWD/temp

# Download commit-analysis tool
wget https://raw.githubusercontent.com/jay/showlinenum/develop/showlinenum.awk
chmod +x showlinenum.awk
mv showlinenum.awk $TMP_DIR

# Generate BBtargets from commit ef709ce2
pushd $SUBJECT
  git checkout ef709ce2
  git diff -U0 HEAD^ HEAD > $TMP_DIR/commit.diff
popd
cat $TMP_DIR/commit.diff |  $TMP_DIR/showlinenum.awk show_header=0 path=1 | grep -e "\.[ch]:[0-9]*:+" -e "\.cpp:[0-9]*:+" -e "\.cc:[0-9]*:+" | cut -d+ -f1 | rev | cut -c2- | rev > $TMP_DIR/BBtargets.txt

# Print extracted targets. 
echo "Targets:"
cat $TMP_DIR/BBtargets.txt

Note: If there are no targets, there is nothing to instrument!

Generate CG and intra-procedural CFGs from the subject.

# Set aflgo-instrumenter
export CC=$AFLGO/instrument/aflgo-clang
export CXX=$AFLGO/instrument/aflgo-clang++

# Set aflgo-instrumentation flags
export COPY_CFLAGS=$CFLAGS
export COPY_CXXFLAGS=$CXXFLAGS
export ADDITIONAL="-targets=$TMP_DIR/BBtargets.txt -outdir=$TMP_DIR -flto -fuse-ld=gold -Wl,-plugin-opt=save-temps"
export CFLAGS="$CFLAGS $ADDITIONAL"
export CXXFLAGS="$CXXFLAGS $ADDITIONAL"

# Build libxml2 (in order to generate CG and CFGs).
# Meanwhile go have a coffee ☕️
export LDFLAGS=-lpthread
pushd $SUBJECT
  ./autogen.sh
  ./configure --disable-shared
  make clean
  make xmllint
popd

You can test whether CG/CFG extraction was successful with

$SUBJECT/xmllint --valid --recover $SUBJECT/test/dtd3
ls $TMP_DIR/dot-files
echo "Function targets"
cat $TMP_DIR/Ftargets.txt

Note:

If the linker (CCLD) complains that you should run ranlib, make sure that libLTO.so and LLVMgold.so (from Install LLVM 11.0.0 with Gold-plugin in step 1) can be found in /usr/lib/bfd-plugins.
If the compiler crashes, there is some problem with LLVM not supporting our instrumentation (afl-llvm-pass.so.cc:540-577). LLVM has changed the instrumentation-API very often :( You can check LLVM-version, fix problem, and prepare pull request.
You can speed up the compilation with a parallel build. However, this may impact which BBs are identified as targets. See aflgo#41.

Generate distance file. Firstly we need to clean up BBnames.txt and BBcalls.txt, otherwise distance_calculator may fail. This is necessary for any subjects, not only for libxml2.

# Clean up
cat $TMP_DIR/BBnames.txt | grep -v "^$"| rev | cut -d: -f2- | rev | sort | uniq > $TMP_DIR/BBnames2.txt && mv $TMP_DIR/BBnames2.txt $TMP_DIR/BBnames.txt

cat $TMP_DIR/BBcalls.txt | grep -Ev "^[^,]*$|^([^,]*,){2,}[^,]*$"| sort | uniq > $TMP_DIR/BBcalls2.txt && mv $TMP_DIR/BBcalls2.txt $TMP_DIR/BBcalls.txt

Then start to generate (this may take a while):

# Generate distance ☕️
# $AFLGO/distance/gen_distance_orig.sh is the original, but significantly slower, version

$AFLGO/distance/gen_distance_fast.py $SUBJECT $TMP_DIR xmllint

After that you can check the generated distance file with

echo "Distance values:"
head -n5 $TMP_DIR/distance.cfg.txt
echo "..."
tail -n5 $TMP_DIR/distance.cfg.txt

Note: If distance.cfg.txt is empty, there was some problem computing the CG-level and BB-level target distance. See $TMP_DIR/step*.log.

Instrument the subject

export CFLAGS="$COPY_CFLAGS -distance=$TMP_DIR/distance.cfg.txt"
export CXXFLAGS="$COPY_CXXFLAGS -distance=$TMP_DIR/distance.cfg.txt"

# Clean and build subject with distance instrumentation ☕️
pushd $SUBJECT
  make clean
  ./configure --disable-shared
  make xmllint
popd

If your compilation crashes in this step, have a look at Issue #4.

How to fuzz the instrumented binary

We set the exponential annealing-based power schedule (-z exp).
We set the time-to-exploitation to 45min (-c 45m), assuming the fuzzer is run for about an hour.

(Still take the previous libxml2 as an example)

# Construct seed corpus
mkdir in
cp -r $SUBJECT/test/dtd* in
cp $SUBJECT/test/dtds/* in

$AFLGO/afl-2.57b/afl-fuzz -S ef709ce2 -z exp -c 45m -i in -o out $SUBJECT/xmllint --valid --recover @@

Tipp: Concurrently fuzz the most recent version as master with classical AFL :)

$AFL/afl-fuzz -M master -i in -o out $MASTER/xmllint --valid --recover @@

Run more fuzzing scripts of various real programs like Binutils, jasper, lrzip, libming and DARPA CGC. Those scripts haven't contained any dependencies installing steps yet. So it's recommended that see READMEs of those projects first to check their requirements.

KilluaYZ/aflgo