Python Security

Python Security model

Python doesn’t implement privilege separation (not “inside” Python) to reduce the attack surface of Python. Once an attacker is able the execute arbitrary Python code, the attacker basically gets a full access to the system. Privilege separation can be implemented “outside” Python by putting Python inside a sandbox.

Example with bpo-36506 (closed as not a bug): getattr() executes arbitrary code by design, it’s not a vulnerability.


CPython doesn’t verify that bytecode is safe. If an attacker is able to execute arbitrary bytecode, we consider that the security of the bytecode is the least important issue: using bytecode, sensitive code can be imported and executed.

For example, the marshal doesn’t validate inputs.


Don’t try to build a sandbox inside CPython. The attack surface is too large. Python has many introspection features, see for example the inspect module. Python also many convenient features which executes code on demand. Examples:

  • the literal string '\N{Snowman}' imports the unicodedata module
  • the code to log a warning might be abused to execute code

The good design is to put CPython into a sandbox, not the opposite.

Ok, understood, but I want a sandbox in Python. Well…

Dangerous functions and modules

  • Python 2 input()
  • Python 2 execfile()
  • eval()
  • subprocess.Popen(shell=True)
  • str.format(), Python 3 str.format_map, and Python 2 unicode.format() all allow arbitrary attribute access on formatted values, and hence access to Python’s introspection features: Be Careful with Python’s New-Style String Format (Armin Ronacher, December 2016)
  • The pickle module executes arbitrary Python code: never use it with untrusted data.

Archives and absolute paths

  • tarfile: Never extract archives from untrusted sources without prior inspection. It is possible that files are created outside of path, e.g. members that have absolute filenames starting with “/” or filenames with two dots “..”.
  • zipfile: Never extract archives from untrusted sources without prior inspection. It is possible that files are created outside of path, e.g. members that have absolute filenames starting with “/” or filenames with two dots “..”. zipfile attempts to prevent that.

Archives and Zip Bomb (CVE-2019-9674)

Be careful of “Zip Bombs”: a very small archive can use a huge amount of memory and disk space once decompressed.

The zlib module allows to limit the maximum length:

For example, the OpenStack Nova was vulnerable of denial of service if a compressed virtual machine was a Zip Bomb: OSSA 2016-012 and CVE-2015-5162.

Turns out qemu image parser is not hardened against malicious input and can be abused to allocated an arbitrary amount of memory and/or dump a lot of information when used with “–output=json”.

Nova has been fixed using the prlimit command (with one implementation written in Python: to limit the maximum memory of the process.


Shell command injection


Whenever possible, avoid subprocess.Popen(shell=True) and os.popen(). On UNIX, shlex.quote() can be used to escape command line arguments to use them safetely in a shell command.

For Windows, see:


The random module must not be used in security sensitive code, except of the random.SystemRandom class.

CPython Security Experts

  • Alex Gaynor
  • Antoine Pitrou
  • Christian Heimes
  • Donald Stufft



ASLR and DEP protections enabled since Python 3.4 (and Python 2.7.11 if built using PCbuild/ directory).

Unsafe Python 2.7 default installation directory

Python 2.7 installer uses C:\Python27\ directory by default. The created directory has the “Modify” access rights given to the “Authenticated Users” group. An attacker can modify the standard library or even modify python.exe. Python 3 installer now installs Python in C:\Program Files by default to fix this issue. Override the default installation directory, or fix the directory permissions.

DLL injection

On Windows 8.1 and older, the installer is vulnerable to DLL injection: evil DLL written in the same download directory that the downloaded Python installer. See DLL Hijacking Just Won’t Die.

DLL injection using PATH

Inject a malicious DLL in a writable directory included in PATH. The “pip” step of the Python installer will run this DLL.

We consider that it is not an issue of Python (Python installer) itself.

Once you have write access to a directory on the system PATH (not the current user PATH) and the ability to write binaries that are not validated by the operating system before loading, there are many more interesting things you can do rather than wait for the Python installer to be run.

Module Search Path (sys.path)

  • python3 -E: ignore PYTHON* environment variables like PYTHONPATH
  • python3 -I: isolated mode, also implies -E and -s
  • bpo-5753: CVE-2008-5983 python: untrusted python modules search path (2009) added PySys_SetArgvEx() (to Python 2.6.6, 2.7.0, 3.1.3, 3.2.0): allows embedders of the interpreter to set sys.argv without also modifying sys.path. This helps fix CVE-2008-5983.
  • CVE-2015-5652: Untrusted search path vulnerability in python.exe in Python through 3.5.0 on Windows allows local users to gain privileges via a Trojan horse readline.pyd file in the current working directory. NOTE: the vendor says “It was determined that this is a longtime behavior of Python that cannot really be altered at this point.”



Python Security Response Team (PSRT)