Need help with polytracker?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

245 Stars 23 Forks Apache License 2.0 1.3K Commits 30 Opened issues


An LLVM-based instrumentation tool for universal taint tracking, dataflow analysis, and tracing.

Services available


Need anything else?

Contributors list


PyPI version Tests Slack Status

PolyTracker is a tool originally created for the Automated Lexical Annotation and Navigation of Parsers, a backronym devised solely for the purpose of referring to it as The ALAN Parsers Project. However, it has evolved into a general purpose tool for efficiently performing data-flow and control-flow analysis of programs. PolyTracker is an LLVM pass that instruments programs to track which bytes of an input file are operated on by which functions. It outputs a database containing the data-flow information, as well as a runtime trace. PolyTracker also provides a Python library for interacting with and analyzing its output, as well as an interactive Python REPL.

PolyTracker can be used in conjunction with PolyFile to automatically determine the semantic purpose of the functions in a parser. It also has an experimental feature capable of generating a context free grammar representing the language accepted by a parser.

Unlike dynamic instrumentation alternatives like Taintgrind, PolyTracker imposes negligible performance overhead for almost all inputs, and is capable of tracking every byte of input at once. PolyTracker started as a fork of the LLVM DataFlowSanitizer and takes much inspiration from the Angora Fuzzer. However, unlike the Angora system, PolyTracker is able to track the entire provenance of a taint. In February of 2021, the LLVM DataFlowSanitizer added a new feature for tracking taint provenance called origin tracking. However, it is only able to track at most 16 taints at once, while PolyTracker can track up to 2³².

This README serves as the general usage guide for installing PolyTracker and compiling/instrumenting binaries. For programmatically interacting with or extending PolyTracker through its Python API, as well as for interacting with runtime traces produced from instrumented code, consult the Python documentation.


PolyTracker is controlled via a Python script called

. You can install it by running
pip3 install polytracker

PolyTracker requires a very particular system environment to run, so almost all users are likely to run it in a virtualized environment. Luckily,

makes this easy. All you need to do is have
installed, then run:
polytracker docker pull


polytracker docker run

The latter command will mount the current working directory into the PolyTracker Docker container, and allow you to build and run instrumented programs.


control script—which you can run from either your host system or from inside the Docker container—has a variety of commands, both for instrumenting programs as well as analyzing the resulting artifacts. For example, you can explore the dataflows in the execution, reconstruct the instrumented program's control flow graph, and even extract a context free grammar matching the inputs accepted by the program. You can explore these commands by running
polytracker --help


script is also a REPL, if run with no command line arguments: ```python $ polytracker PolyTracker (3.0.0) Type "help" or "commands"

commands ```

Instrumenting a simple C/C++ program

Installing PolyTracker will also install two build scripts:

. These scripts are essentially wrappers around
and have similar arguments. In the Docker container, these are mapped to
. If run from the host system, these scripts will automatically and seamlessly perform the build within Docker, if necessary.

If you have a C target, you can instrument it by invoking

and passing the
before your cflags:
polybuild --instrument-target -g -o my_target my_target.c 

Repeat the same steps above for a cxx file by invoking

instead of

For more complex programs that use a build system like autotools or CMake, or generally for programs that have multiple compilation units, ensure that the build program uses

(e.g., by setting the
environment variable), and compile the program as normal:
$ CC=polybuild make
Then run this on the resulting binary:
$ get-bc -b the_binary
$ polybuild --lower-bitcode -i the_binary.bc -o the_binary_polytracker --libs LIST_OF_LIBRARIES_TO_LINK
will be the instrumented version. See the Dockerfiles in the examples directory for examples of how real-world programs can be instrumented.

Running and Analyzing an Instrumented Program

The PolyTracker instrumentation looks for the

environment variable to specify which input file's bytes are meant to be tracked. (Note: PolyTracker can in fact track multiple input files—and really any file-like stream such as network sockets—however, we have thus far only exposed the capability to specify a single file. This will be improved in a future release.)

The instrumented software will write its output to the path specified in

, or
if omitted. This is a sqlite3 database that can be operated on by running: ```python from polytracker import PolyTrackerTrace

trace = PolyTrackerTrace.load("polytracker.db")

for event in trace: # this prints every single event in the trace, in order print(event)

for function in trace.functions: # this returns a function object for every function observed # (this is different than a function invocation, which we describe below) print(function.demangled_name)

mainfunc = trace.getfunction("main") for taint in main_func.taints().regions(): # this iterates over every taint touched by the main function aggregated across all invocations of main print(f"source={taint.source}, offset={taint.offset}, length={taint.length}, value={taint.value}")

Individual function invocations in the trace can also be enumerated:
python entrypoint = trace.entrypoint

entrypoint is a function invocation object associated with a single invocation of the entrypoint of the trace (main)

print(f"Entrypoint is: {entrypoint!s}") for calledfunction in entrypoint.calls(): # calledfunction is another function invocation object # associated with a function called from entrypoint print(str(calledfunction)) for region in calledfunction.taints().regions(): # this will iterate over every tainted region operated on by calledfunction, # including any functions called to from calledfunction print(f"\t{region.value}") ```

You can also run an instrumented binary directly from the REPL: ```python $ polytracker PolyTracker (3.0.0) Type "help" or "commands"

trace = runtrace("pathtobinary", "pathtoinputfile") for event in trace: ... print(event) ``` This will automatically run the instrumented binary in a Docker container, if necessary.

:warning: If running PolyTracker inside Docker or a VM: PolyTracker can be very slow if running in a virtualized environment and either the input file or, especially, the output database are located in a directory mapped or mounted from the host OS. This is particularly true when running PolyTracker in Docker from a macOS host. The solution is to write the database to a path inside of the container/VM and then copy it out to the host system at the very end.

The optional

environment variable can be set to
to produce a basic-block level trace of the program.

The Python API documentation is available here.

Runtime Parameters and Instrumentation Tuning

At runtime, PolyTracker instrumentation looks for a number of configuration parameters either specified through environment variables or a local configuration file. This allows one to modify instrumentation parameters without needing to recompile the binary.

Environment Variables

PolyTracker accepts configuration parameters in the form of environment variables to avoid recompiling target programs. The current environment variables PolyTracker supports is:

POLYPATH: The path to the file to mark as tainted 

POLYTTL: This value is an initial "strength" value for taint nodes, when new nodes are formed, the average is taken. When the TTL value is 0, the node is considered clean.

POLYSTART: Start offset to track

POLYEND: End offset to track

POLYDB: A path to which to save the output database (default is polytracker.db)

POLYCONFIG: Provides a path to a JSON file specifying settings

WLLVM_ARTIFACT_STORE: Provides a path to an existing directory to store artifact/manifest for all build targets

Configuration Files

Rather than setting environment variables on every run, you can make a configuration file.


    "POLYSTART": 1,
    "POLYEND": 3,
    "POLYTTL": 16

Polytracker will set its configuration parameters in the following order: 1. If a parameter is specified via an environment variable, use that value 2. Else if

is specified and that configuration file contains the parameter, use that value 3. Else if the current directory contains
and that config contains the parameter, use that value 4. Else if
exists and it contains the parameter, use that value 5. Else if a default value for the parameter exists, use the default 6. Else throw an error

ABI Lists

DFSan uses ABI lists to determine what functions it should automatically instrument, what functions it should ignore, and what custom function wrappers exist. See the dfsan documentation for more information.

Creating custom ignore lists from pre-built libraries

Attempting to build large software projects can be time consuming, especially older/unsupported ones. It's even more time consuming to try and modify the build system such that it supports changes, like dfsan's/our instrumentation.

There is a script located in

that you can run on any ELF library and it will output a list of functions to ignore. We use this when we do not want to track information going through a specific library like libpng, or other sub components of a program. The
exists to build common open source libraries so we can create these lists.

This script is a slightly tweaked version of what DataFlowSanitizer has, which focuses on ignoring system libraries. The original script can be found in


Building the Examples

Check out this Git repository. From the root, either build the base PolyTracker Docker image:

pip3 install -e .[dev] && polytracker docker rebuild

or pull the latest prebuilt version from DockerHub:

docker pull trailofbits/polytracker:latest

This will create a Docker container with PolyTracker built, and the

environment variable set to
. Simply add the code to be instrumented to this container, and as long as its build process honors the
environment variable, the resulting binary will be instrumented.

For a demo of PolyTracker running on the MuPDF parser run this command:

docker build -t trailofbits/polytracker-demo-mupdf -f examples/pdf/Dockerfile-mupdf.demo .

will be build in
. Running
will output
which contains the information provided by the taint analysis. Its recommended to use this json with PolyFile.

For a demo of PolyTracker running on Poppler utils version 0.84.0 run this command:

docker build -t trailofbits/polytracker-demo-poppler -f examples/pdf/Dockerfile-poppler.demo .

All the poppler utils will be located in

$ cd /polytracker/the_klondike/poppler-0.84.0/build/utils
$ POLYPATH=some_pdf.pdf ./pdfinfo_track some_pdf.pdf

Building PolyTracker from Source

The compilation process for both PolyTracker LLVM and PolyTracker is rather fickle, since it involves juggling both instrumented and non-instrumented versions of standard library bitcode. We highly recommend using our pre-built and tested Docker container if at all possible. Installing the PolyTracker Python package on your host system will allow you to seamlessly interact with the prebuilt Docker container. Otherwise, to install PolyTracker natively, we recommend first replicating the install process from the

Dockerfile, followed by replicating the install process from the PolyTracker Dockerfile.

Build Dependencies

  • PolyTracker LLVM. PolyTracker is built atop its own fork of LLVM,
    . This fork modifies the DataFlow Sanitizer to use increased label sizes (to allow for tracking orders of magnitude more taints), as well as alternative data structures to store them. We have investigated up-streaming our changes into LLVM proper, but there has been little interest. The changes are relatively minor, so keeping the fork in sync with upstream LLVM should be relatively straightforward.
  • CMake
  • Ninja (
    on Ubuntu)

Runtime Dependencies

The following tools are required to test and run PolyTracker: * Python 3.7+ and

apt-get -y install python3.7 python3-pip
). These are used for both seamlessly interacting with the Docker container (if necessary), as well as post-processing and analyzing the artifacts produced from runtime traces. * gllvm (
go get
) is used to create whole program bitcode archives and to extract bitcode from targets.

Building with LLVM Xray Instrumentation

It's possible to build both Polytracker and target programs with LLVM's Xray instrumentation. Pass the

option to Polytracker, and build targets with
to build targets with xray instrumentation.

An example on how to use Xray lives here.

Current Status and Known Issues

PolyTracker currently only runs on Linux, because that is the only system supported by the DataFlow Santizer. This limitation is just due to a lack of support for semantics for other OSes system calls, which could be added in the future. However, this means that running PolyTracker on a non-Linux system will require Docker to be installed.

Taints will not propagate through dynamically loaded libraries unless those libraries were compiled from source using PolyTracker, or there is specific support for the library calls implemented in PolyTracker. There is currently support for propagating taint through the majority of uninstrumented C standard library calls. To be clear, programs that use uninstrumented functions will still run normally, however, operations performed by unsupported library calls will not propagate taint. We are currently working on adding robust support for C++ programs, but currently the best results will be from C programs.

If there are issues with Docker, try performing a system prune and build with

for both PolyTracker and whatever demo you are trying to run.

The worst case performance of PolyTracker is exercised when a single byte in memory is simultaneously tainted by a large number of input bytes from the source file. This is most common when instrumenting compression and cryptographic algorithms that have large block sizes. There are a number of mitigations for this behavior currently being researched and developed.

License and Acknowledgements

This research was developed by Trail of Bits with funding from the Defense Advanced Research Projects Agency (DARPA) under the SafeDocs program as a subcontractor to Galois. It is licensed under the Apache 2.0 license. © 2019, Trail of Bits.


Carson Harmon
Evan Sultanik
Brad Larsen

[email protected]

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.