Python Security Manager

What is Python Security Manager? {#what}

The Python Security Manager (pysecmgr) is a middleware security layer in Python that can monitor and prevent insecure logic from being run.

Note: pysecmgr is not a sandbox, see #limits for more details.

Which python version is supported?

Pysecmgr needs at least python version 3.8, as it first introduced PEP 0578. Pysecmgr has been tested using python 3.9.

Why should I use it? {#why}

Pysecmgr is a defense in depth mechanism to protect against exploitation of certain attacks including:

• Path traversal and arbitrary files disclosure
• OS command injection
• Code injection
• Arbitrary outbound network connections (including SSRF)
• Source code disclosure
• Insecure deserialization

To better understand how pysecmgr can help protect against those attacks, see the pysecmgr codelab in pysecmgr/codelab/README.md

When to use pysecmgr?

• Web applications are one of the primary use cases for pysecmgr as they usually handle untrusted user inputs.
• Pysecmgr can be used by individual applications, and can also be integrated into frameworks to protect a wider audience.

When not to use pysecmgr?

• For tooling related scripts that don’t take untrusted user inputs, pysecmgr might be too restrictive (e.g. unable to execute system command, restricted popen args).
• Pysecmgr is not a sandbox and should not be used as the only security control to secure/restrict arbitrary and untrusted code provided by users. See pysecmgr limitations.

How does it work? {#how}

Pysecmgr relies on PEP 0578 which introduces audit hooks allowing security-sensitive Python methods to be inspected by callback methods. During inspection, exceptions can be raised and this allows the callback to prevent a method from being called/executed.

The Python Security Manager currently supports the following hooks:

• open: pysecmgr hooks the "open" built-in function, os.open and io.open.
• socket.connect: pysecmgr restrict socket ip and port or UNIX socket path.
• pickle.find_class: pysecmgr prevents unpickling event by default.
• os.system: pysecmgr prevents the use of os.system by default.
• subprocess.Popen: pysecmgr restricts the arguments that can be passed to popen.
• exec: pysecmgr hooks the "exec" and "eval" build-in function. Note that this hook’s allowlist behaves differently than the others as it does not allowlist the content being passed to exec but the Python modules that can use exec.
• import: this hook is currently not enabled by default because it requires a lot of customizations for each application. It is considered an advanced feature for more security sensitive use cases.

The following table describes other pysecmgr supported functions whose implementation is based on the above hooked functions. This is not a holistic list and will be improved over time.

Functions	Support
eval	Supported, uses os.exec.
os.popen	Supported, uses subprocess.Popen.
subprocess.run	Supported, uses subprocess.Popen.
os.spawn*	Supported, uses os.exec.
urllib.request.urlopen	Supported, uses socket.connect.
os.exec*	Supported, uses os.exec.

Limitations {#limits}

Pysecmgr has a few known limitations:

• Pysecmgr is not catching bugs introduced by C/C++ (SWIG) code.
• Some common python vulnerabilities are not covered by pysecmgr such as insecure yaml or XXE with lxml. Pyyaml is not using a pysecmgr supported audit hook to create python objects. Similarly, lxml also does not use supported audit hook.
• Pysecmgr is not a sandbox, it does not hook all security sensitive calls (incl. C/C++ (SWIG) code) and can be bypassed.
• Pysecmgr does not patch insecure logic. It only blocks them.
• Some functions such as shutil.rmtree are using the low level os.open with a path relative to a directory file descriptor (dir_fd). This is unfortunately not supported by pysecmgr.

Disclaimer

This is not an officially supported Google product.