Skip to main content

Shipping Pysa Models and Rules

We welcome contributions and coverage improvements for Pysa! This page walks you through the basics of adding new models or rules, and testing your changes.


This document assumes you:

  1. Have followed these instructions to set up Pysa and SAPP.
  2. Understand the basic concepts of sources, sinks, sanitizers, and rules as well as features.
  3. Have completed a successful run of Pysa
  4. Know the sources and sinks you'd like to detect.

Ideally, you will also have completed the Pysa tutorial (code, video).

Where to put your models / taint.config changes#

Common privacy/security-related rules and models should be placed under stubs/taint/core_privacy_security. These should only reference entities that exist in the Python standard library, or the type stubs that are shipped with Pyre.

Models for third-party libraries should be placed under stubs/third_party_taint.

Writing your models / taint.config changes#

Refer to the docs on sources, sinks, sanitizers, rules, and features when you need help. You can usually find pre-existing examples of sources, sinks, sanitizers, and features to copy as a starting point if you get stuck.

When adding models for a pre-existing source/sink/feature type (e.g. UserControlled, you won't need to modify taint.config. Make sure to check for pre-existing source, sink, and feature declarations before declaring a new one; most types of source and sinks are already declared, and it can be more sustainable to add an additional model to an existing category, rather than creating a whole new one.

When adding a new rule, you will need to update taint.config. As with the previous paragraph, try to put existing sources and sinks to use. For example, if you're writing a sensitive data logging rule, using the Logging sink will allow you to detect your chosen data flowing into many different types of loggers that we model.


Providing a quick explanation of the issue you intended to catch and evidence that the issue was caught in your local run is usually sufficient. This can be in the form of a screenshot of the issue in SAPP, a sample or paste of the taint-output.json produced by Pysa, etc.

Catching Known Issues#

To test, you need to have an issue that you want to find. The best option is to have a known vulnerability from a past CVE or issue. You can manually identify the flow of data that caused the issue, make sure you have the correct sources and sinks, and then verify that your new rule/source/sink catches the issue.

Follow these instructions to run Pysa and import the results into the SAPP UI.

Once your run completes, you should be able to see the issue you intended to catch in the UI. If there are a lot of issues showing up, you may need to use filters to find the particular issue you were looking for. If you're not finding your issue, read through the development tips for help debugging.

Integration Test#

If you don't have an existing project to test on, you can also use the integration test environment provided with the pyre-check repo. This is a minimal, deliberately vulnerable Flask web app. You can find all the details and instructions on how to set it up here.

It should already be set up to use the taint models in the pyre-check/stubs folder, so you can easily make your changes to existing Pysa models and observe the effects.

For example, suppose we wanted to add a new rule for some builtin Python functions. We would add our new source, sink and rule in stubs/taint/core_privacy_security/taint.config:

{    ...    "sources": [        ...,        {            "name": "Foo",            "comment": "used to annotate a foo source"        },    ],    "sinks": [        ...,        {            "name": "Bar",            "comment": "used to annotate a bar sink"        },    ],    "rules": [        ...,        {            "name": "Foo to Bar",            "code": 9000,            "sources": [                "Foo"            ],            "sinks": [                "Bar"            ],            "message_format": "Data from [{$sources}] source(s) may reach [{$sinks}] sink(s)"        },    ],}

Make sure your new rule has a unique number and new source and sink names are also unique!

We can now add our models in a .pysa file under stubs/taint/core_privacy_security/:

def input() -> TaintSource[Foo]: ...def ascii(__source: TaintSink[Bar]): ...

Then, we can open a source file in the vulnerable Flask app (e.g., and inject an issue of the type we want to catch:

def alarm1() -> None:    x = input()    ascii(x)

Finally, we can run the integration test using the script (or just run pyre analyze) and verify that the issue we expect to be caught is indeed caught:

ERROR ----BEGIN PYSA INTEGRATION TEST ERROR----ERROR Output differs from expected:...@@ -46,5 +46,11 @@    ...+  },+  {+    "code": 9000,+    "define": "app.alarm1",+    "description": "Foo to Bar [9000]: Data from [Foo] source(s) may reach [Bar] sink(s)",+    "path": ""   } ]ERROR ----END PYSA INTEGRATION TEST ERROR----

Contributing Coverage Improvements to Pysa#

When you've proven that you can catch issues with your changes, send a PR to the pyre-check Github repository. Please make sure you include a test plan in your PR that follows the testing guidelines mentioned above!

If you used the deliberately_vulnerable_flask_app for testing, free to also include your integration test changes with your injected issue in your PR to help us expand our open source integration tests! Note this will require you to update the full_result.json file in the same folder. When you run the integration test script, a full_result.actual file will be produced when the output does not match the existing expected output. Simply copy the contents of that file to full_result.json, e.g. in the example above, we would add the following lines to full_result.json:

    ...    {        "code": 9000,        "column": 10,        "define": "app.alarm1",        "description": "Foo to Bar [9000]: Data from [Foo] source(s) may reach [Bar] sink(s)",        "line": 64,        "name": "Foo to Bar",        "path": "",        "stop_column": 11,        "stop_line": 64    }]