Fuzzilli

A (coverage-)guided fuzzer for dynamic language interpreters based on a custom intermediate language ("FuzzIL") which can be mutated and translated to JavaScript.

Usage

The basic steps to use this fuzzer are:

1. Download the source code for one of the supported JavaScript engines. See the Targets/ directory for the list of supported JavaScript engines.
2. Apply the corresponding patches from the target's directory. Also see the README.md in that directory.
3. Compile the engine with coverage instrumentation (requires clang >= 4.0) as described in the README.
4. Compile the fuzzer: swift build [-c release].
5. Run the fuzzer: swift run [-c release] FuzzilliCli --profile=<profile> [other cli options] /path/to/jsshell. See also swift run FuzzilliCli --help.

Building and running Fuzzilli and the supported JavaScript engines inside Docker and on Google Compute Engine is also supported.

Hacking

Check out main.swift to see a usage example of the Fuzzilli library and play with the various configuration options. Next, take a look at Fuzzer.swift for the highlevel fuzzing logic. From there dive into any part that seems interesting.

Patches, additions, other contributions etc. to this project are very welcome! However, do quickly check the notes for contributors. Fuzzilli roughly follows Google's code style guide for swift.

It would be much appreciated if you could send a short note (possibly including a CVE number) to saelo@google.com or open a pull request for any vulnerability found with the help of this project so it can be included in the bug showcase section. Other than that you can of course claim any bug bounty, CVE credits, etc. for the vulnerabilities :)

Concept

When fuzzing for core interpreter bugs, e.g. in JIT compilers, semantic correctness of generated programs becomes a concern. This is in contrast to most other scenarios, e.g. fuzzing of runtime APIs, in which case semantic correctness can easily be worked around by wrapping the generated code in try-catch constructs. There are different possibilities to achieve an acceptable rate of semantically correct samples, one of them being a mutational approach in which all samples in the corpus are also semantically valid. In that case, each mutation only has a small chance of turning a valid sample into an invalid one.

To implement a mutation-based JavaScript fuzzer, mutations to JavaScript code have to be defined. Instead of mutating the AST, or other syntactic elements of a program, a custom intermediate language (IL) is defined on which mutations to the control and data flow of a program can more directly be performed. This IL is afterwards translated to JavaScript for execution. The intermediate language looks roughly as follows:

v0 <− LoadInteger '0'
v1 <− LoadInteger '10'
v2 <− LoadInteger '1'
v3 <− LoadInteger '0'
BeginFor v0, '<', v1, '+', v2 −> v4
   v6 <− BinaryOperation v3, '+', v4
   Reassign v3, v6
EndFor
v7 <− LoadString 'Result: '
v8 <− BinaryOperation v7, '+', v3
v9 <− LoadGlobal 'console'
v10 <− CallMethod v9, 'log', [v8]

Which can e.g. be trivially translated to the following JavaScript code:

const v0 = 0;
const v1 = 10;
const v2 = 1;
let v3 = 0;
for (let v4 = v0; v4 < v1; v4 = v4 + v2) {
    const v6 = v3 + v4;
    v3 = v6;
}
const v7 = "Result: ";
const v8 = v7 + v3;
const v9 = console;
const v10 = v9.log(v8);

Or to the following JavaScript code by inlining intermediate expressions:

let v3 = 0;
for (let v4 = 0; v4 < 10; v4++) {
    v3 = v3 + v4;
}
console.log("Result: " + v3);

FuzzIL has a number of properties:

• A FuzzIL program is simply a list of instructions.
• A FuzzIL instruction is an operation together with input and output variables and potentially one or more parameters (enclosed in single quotes in the notation above).
• Inputs to instructions are always variables, there are no immediate values.
• Every output of an instruction is a new variable, and existing variables can only be reassigned through dedicated operations such as the Reassign instruction.
• Every variable is defined before it is used.

A number of mutations can then be performed on these programs:

• InputMutator: replaces input variables of instructions with different ones to mutate the dataflow of the program.
• CodeGenMutator: generates code and inserts it somewhere in the mutated program. Code is generated either by running a code generator or by copying some instructions from another program in the corpus (splicing).
• CombineMutator: inserts a program from the corpus into a random position in the mutated program.
• OperationMutator: mutates the parameters of operations, for example replacing an integer constant with a different one.
• and more...

A much more thorough discussion of how Fuzzilli works can be found here.

Implementation

The fuzzer is implemented in Swift, with some parts (e.g. coverage measurements, socket interactions, etc.) implemented in C.

Architecture

A fuzzer instance (implemented in Fuzzer.swift) is made up of the following central components:

• MutationFuzzer: produces new programs from existing ones by applying mutations. Afterwards executes the produced samples and evaluates them.
• ScriptRunner: executes programs of the target language.
• Corpus: stores interesting samples and supplies them to the core fuzzer.
• Environment: has knowledge of the runtime environment, e.g. the available builtins, property names, and methods.
• Minimizer: minimizes crashing and interesting programs.
• Evaluator: evaluates whether a sample is interesting according to some metric, e.g. code coverage.
• Lifter: translates a FuzzIL program to the target language (JavaScript).

Furthermore, a number of modules are optionally available:

• Statistics: gathers various pieces of statistical information.
• NetworkSync: synchronize multiple instances over the network.
• ThreadSync: synchronize multiple instances within the same process.
• Storage: stores crashing programs to disk.

The fuzzer is event-driven, with most of the interactions between different classes happening through events. Events are dispatched e.g. as a result of a crash or an interesting program being found, a new program being executed, a log message being generated and so on. See Events.swift for the full list of events. The event mechanism effectively decouples the various components of the fuzzer and makes it easy to implement additional modules.

A FuzzIL program can be built up using a ProgramBuilder instance. A ProgramBuilder provides methods to create and append new instructions, append instructions from another program, retrieve existing variables, query the execution context at the current position (e.g. whether it is inside a loop), and more.

Execution

Fuzzilli uses a custom execution mode called REPRL (read-eval-print-reset-loop). For that, the target engine is modified to accept a script input over pipes and/or shared memory, execute it, then reset its internal state and wait for the next script. This removes the overhead from process creation and to a large part from the engine ininitializaiton.

Scalability

There is one Fuzzer instance per target process. This enables synchronous execution of programs and thereby simplifies the implementation of various algorithms such as consecutive mutations and minimization. Moreover, it avoids the need to implement thread-safe access to internal state, e.g. the corpus. Each fuzzer instance has its own DispatchQueue, conceptually corresponding to a single thread. As a rule of thumb, every interaction with a Fuzzer instance must happen on that instance’s dispatch queue. This guarantees thread-safety as the queue is serial. For more details see the docs.

To scale, fuzzer instances can form a tree hierarchy, in which case they report newly found interesting samples and crashes to their parent node. In turn, a parent node synchronizes its corpus with its child nodes. Communication between nodes in the tree can happen in different ways, each implemented as a module:

• Inter-thread communication: synchronize instances in the same process by enqueuing tasks to the other fuzzer’s DispatchQueue.
• Inter-machine communication: synchronize instances over a simple TCP-based protocol.

This design allows the fuzzer to scale to many cores on a single machine as well as to many different machines. As one parent node can quickly become overloaded if too many instances send programs to it, it is possible to configure multiple levels of instances, e.g. one root instance, 16 intermediate nodes connected to the root, and 256 "leaves" connected to the intermediate nodes. See the Cloud/ directory for more information about distributed fuzzing.

Resources

Further resources about this fuzzer:

• A presentation about Fuzzilli given at Offensive Con 2019.
• The master's thesis for which the initial implementation was done.
• A blogpost by Sensepost about using Fuzzilli to find a bug in v8.
• A blogpost by Doyensec about fuzzing the JerryScript engine with Fuzzilli.
• A paper from the NDSS Symposium 2023 about Fuzzilli and how it compares to other fuzzers.

Bug Showcase

The following is a list of some of the bugs found with the help of Fuzzilli. Only bugs with security impact that were present in at least a Beta release of the affected software should be included in this list. Since Fuzzilli is often used for continuous fuzz testing during development, many issues found by it are not included in this list as they are typically found prior to the vulnerable code reaching a Beta release. A list of all issues recently found by Fuzzilli in V8 can, however, be found here.

Special thanks to all users of Fuzzilli who have reported bugs found by it!

WebKit/JavaScriptCore

• Issue 185328: DFG Compiler uses incorrect output register for NumberIsInteger operation
• CVE-2018-4299: performProxyCall leaks internal object to script
• CVE-2018-4359: compileMathIC produces incorrect machine code
• CVE-2019-8518: OOB access in FTL JIT due to LICM moving array access before the bounds check
• CVE-2019-8558: CodeBlock UaF due to dangling Watchpoints
• CVE-2019-8611: AIR optimization incorrectly removes assignment to register
• CVE-2019-8623: Loop-invariant code motion (LICM) in DFG JIT leaves stack variable uninitialized
• CVE-2019-8622: DFG's doesGC() is incorrect about the HasIndexedProperty operation's behaviour on StringObjects
• CVE-2019-8671: DFG: Loop-invariant code motion (LICM) leaves object property access unguarded
• CVE-2019-8672: JSValue use-after-free in ValueProfiles
• CVE-2019-8678: JSC fails to run haveABadTime() when some prototypes are modified, leading to type confusions
• CVE-2019-8685: JSPropertyNameEnumerator uses wrong structure IDs
• CVE-2019-8765: GetterSetter type confusion during DFG compilation
• CVE-2019-8820: Type confusion during bailout when reconstructing arguments objects
• CVE-2019-8844: ObjectAllocationSinkingPhase shouldn't insert hints for allocations which are no longer valid
• CVE-2020-3901: GetterSetter type confusion in FTL JIT code (due to not always safe LICM)
• CVE-2021-30851: Missing lock during concurrent HashTable lookup
• CVE-2021-30818: Type confusion when reconstructing arguments on DFG OSR Exit
• CVE-2022-46696: Assertion failure due to missing exception check in JIT-compiled code
• CVE-2022-46699: Assertion failure due to incorrect caching of special properties in ICs
• CVE-2022-46700: Intl.Locale.prototype.hourCycles leaks empty JSValue to script
• CVE-2025-43214: Memory corruption during JSToWasmEntry when iterating over the stack
• CVE-2025-43213: Invalid typing of NewRegExpUntyped operation

Gecko/Spidermonkey

• [CVE-2018-12386](https://ssd-disclosure.com/archives/3765/ssd-advisory-firefox