Skip to content
Go to file


Failed to load latest commit information.
Latest commit message
Commit time

This is a fork server based approach to fuzz Java applications on Java virtual machine with american fuzzy lop. See caveats section about the downsides of this approach.


Fuzzing with american fuzzy lop works by instrumenting the compiled Java bytecode with probabilistic program coverage revealing instrumentation. There are general types of instrumeting fuzzing modes in programs that can be fuzzed with afl-fuzz command. The default fork server mode does not need any modifications to the program source code and can work as is. There are also more efficient deferred fork server and persistent modes that enable you to skip some initialization code and keep the JVM running longer than for just one input.

Ahead of time instrumentation

Ahead of time instrumentation works by instrumenting specific .jar or .class files that you want to run with afl-fuzz for your program. This is done by running the built java-afl-instrument.jar and instrumenting each jar or class file that you want to include in your program. No source code modifications are necessary to get started:

$ java -jar java-afl-instrument.jar instrumented/ ClassToTest.class
$ java -jar java-afl-instrument.jar instrumented/ jar-to-test.jar

As instrumentation injects native JNI code into the used files, so you can only run these files on similar enough systems that java-afl-instrument.jar was run on.

Then you are ready to fuzz your Java application with afl-fuzz. It can be done with this type of command with the provided java-afl-fuzz wrapper script:

$ java-afl-fuzz -m 20000 -i in/ -o /dev/shm/fuzz-out/ -- java -cp instrumented/ ClassToTest
$ java-afl-fuzz -m 20000 -i in/ -o /dev/shm/fuzz-out/ -- java -jar instrumented/jar-to-test.jar

Just in time instrumentation

Just in time instrumentation works by wrapping the main function of a program that you want to run around a custom instrumentation injecting ClassLoader. This way you will get more thorough instrumentation than just running ahead of time instrumentation on your program, but at the same time the instrumentation likely covers code that you are not interested in.

Just in time instrumentation works by adding both java-afl-run.jar and the target classes to CLASSPATH and running class with the target class name as a parameter:

$ java-afl-fuzz -m 20000 -i in/ -o /dev/shm/fuzz-out/ \
      -- java -cp java-afl-run.jar:. ClassToTest
$ java-afl-fuzz -m 20000 -i in/ -o /dev/shm/fuzz-out/ \
      -- java -cp java-afl-run.jar:jar-to-test.jar ClassToTest

Notice that there is no need to first instrument the class files, as it is done on fly. This has the same platform specific limitations as ahead of time compilation, as this instrumentation injects native JNI code into the used files. So you can only fuzz programs with java-afl-run.jar on similar enough systems that java-afl-run.jar was built on.

java-afl-fuzz parameters

Parameters to java-afl-fuzz command have following functions:

  • -i in/: Input directory of initial data that then gets modified over the fuzzing process.
  • -o /dev/shm/fuzz-out/: Output directory for fuzzing state data. This should always be on a shared memory drive and never in a directory pointing to a physical hard drive.
  • -m 20000: Higher virtual memory limit that enables JVM to run, as the default memory limit in afl-fuzz is 50 megabytes. JVM can allocate around 10 gigabytes of virtual memory by default.

More detailed description of available options can be found from american fuzzy lop's README. You may also want to adjust maximum heap size with -Xmx option to be smaller than the default if you fuzz multiple JVM instances on the same machine to keep memory usage sane.

Advanced usage

More efficient deferred and persistent modes start each fuzzing iteration later than at the beginning of main() function. Using deferred or persistent mode requires either a special annotation for the main() function or --custom-init flag to the instrument program:

public class ProgramCustom {
    public static void main(String args[]) {

Or you can instrument unmodified code in such way that the init function does not need to reside inside main() by making --custom-init as the first parameter:

$ java -jar java-afl-instrument.jar --custom-init instrumented/ ClassToTest.class
$ java -jar java-afl-instrument.jar --custom-init instrumented/ jar-to-test.jar

To put the application into deferred mode where all the initialization code that comes before javafl.fuzz.init() function can be done in following fashion:

public class ProgramPersistent {
    public static void main(String[] args) {
        // You need to read the actual input after initialization point.;
        ... do actual input processing...

To put the program into a persistent mode you need wrap the part that you want to execute around a while (javafl.fuzz.loop(<iterations>)) loop. If you read the input from, you need to take care that you flush Java's buffering on it after you have read your data:

public class ProgramPersistent {
    public static void main(String[] args) {
        byte[] data = new byte[128];
        int read = 128;
        while (javafl.fuzz.loop(100000)) {
            read =, 0, data.length);
            // Throw away all buffering information from stdin for the
            // next iteration:
            ... do actual input processing...

Options controlling instrumentation

Command line switches to java-afl-instrument.jar:

  • --custom-init
  • --deterministic: by default java-afl produces random class files to make it possible to probabilistically get bigger coverage on the program from two differently instrumented programs than from one. This switch makes the instrumentation depend solely on the input data for each class and will always result in the same result between different instrumentation runs. Just in time instrumentation is always deterministic.

Environmental variables:

  • AFL_INST_RATIO: by default 100% of program control flow altering locations are instrumented. This makes it possible to probabilistically select a smaller instrumentation ratio. Smaller instrumentation ratios are useful in big programs where resulting program execution path traces would otherwise fill the default 16 bit state map and increasing the map size would add unneeded performance penalty.


As there are tons of different tools to build Java programs with automatic dependency fetching, java-afl supports more than one way to build itself.

If you pass american fuzzy lop's source code directory that has config.h file in it, you can pass following C flags to JNI compilation part:

CFLAGS="-I<path-to-afl-src-dir> -DHAVE_AFL_CONFIG_H"

This makes the compiled information match to what afl-fuzz expects if it has been modified in any way. Build systems also try to deduce this (TODO) during compilation from existing afl-showmap command if such exists.


Bazel a build tool that can handle very large programs with ease.

$ bazel build :java-afl-instrument_deploy.jar :java-afl-run_deploy.jar
# Stand-alone jars are under  bazel-bin/ as java-afl-instrument_deploy.jar and java-afl-run_deploy.jar


CMake is the PHP of build systems. Widely available and gets stuff done but becomes quite painful after a while.

$ ( mkdir -p build-cmake && cd build-cmake && cmake .. -GNinja )
$ ninja -C build-cmake
# Stand-alone jars are under build-cmake/ as java-afl-instrument.jar and java-afl-run.jar

Travis CI Build Status

Requires Ubuntu 14.04 based system. You need to have ASM 6.1 to build this as a dependency in addition to Java 8 and afl build dependencies. Currently there is a crude build script to build and test this implementation:

$ ./

Even though building requires Java 8, this should be able to instrument programs that run only on some older versions of Java.


Performance numbers on Intel Core i7-3770K CPU @ 3.50GHz with OpenJDK 1.8.0_151 and afl 2.52b. These tests were done with the simple test programs that are provided at test/ directory.

  • Fork server mode around 750 executions/second for a program that does nothing. Closer to 300 when there is actually something happening.
  • Deferred mode naturally gets something between the fork server mode and persistent mode. Depends how heavy the initialization is, probably maybe some tens of percents.
  • Persistent mode around 14000 executions/second. Highly depends on how much and how long JVM is able to optimize before being killed. See caveats section about this. Around 31000 iterations/second for an empty while loop, that is close to the maximum that native C code can handle with afl-fuzz in persistent mode.


  • Fix persistent mode loop dynamic instrumentation.
  • Check if a dynamically instrumentable class is a file and load it or a full jar file instead.
  • Support deferred init for arbitrary given method without source code modifications. Just prefer the loop syntax and non-forking mode instead of fork server one for more speed.
    • Remove the need for @javafl.CustomInit.
  • Create a non-forking alternative mode.
  • More ways to build this:
    • Ant
    • Maven
    • Gradle
  • Alternative method implementations based on fuzzing mode (similar to C preprocessor's #ifdef/#ifndef). Probably somehow with annotations or System.getProperty("FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION").


  • Great thanks to Michał Zalewski for american fuzzy lop, a crude but effective fuzzer. Especially the idea of using a bitmap and randomly generated program locations as a fast probabilistic memory bound approximation of the program execution path.
  • Inspired by python-afl and Kelinci. Just in time instrumentation idea from JQF.


Mandatory dependencies to build this:

  • GNU/Linux system with recently new basic utilities.
  • C compiler.
  • Java 1.8 or newer to build and to instrument classes with afl-fuzz compatible instrumentation. Runtime Java version of instrumented classes should be anything that the original class worked with.
  • ASM 6.1

Optional dependencies for building include one of these:

  • Bazel
  • CMake


Java virtual machine is a multi-threaded application and fork() call only preserves the thread that called it. This creates some issues from performance and stability point of view:

  • There is no garbage collector running in the forked process (at least with OpenJDK). Therefore there is a limit on objects that the fuzz target can allocate during its lifetime. This shouldn't be an issue with a generally lightweight fuzz targets that can execute hundreds of times per second, but can become an issue with more heavy ones.
  • Performance will suffer, as JVM will not be able to use knowledge about hotspots in often executed functions.
  • Persistent mode has a limited number of cycles that it can run before it runs out of memory due to no garbage collector running. TODO create a non-forking alternative for persistent mode.
    • This will make afl-fuzz to result in a timeout every so often when the program runs out of some resource. If the timeout is set manually to be relatively long in otherwise fast fuzz target, it will needlessly delay the recovery from a resource leaking situation.


Copyright 2018 Jussi Judin

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

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.


Binary rewriting approach with fork server support to fuzz Java applications with afl-fuzz.




No releases published


No packages published
You can’t perform that action at this time.