README

!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
!!! WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING !!!
!!!                                                                 !!!
!!! pysandbox is BROKEN BY DESIGN, please move to a new sandboxing  !!!
!!! solution: run python in a sandbox, not the opposite!            !!!
!!!                                                                 !!!
!!! Learn more about pysandbox failure:                             !!!
!!! https://lwn.net/Articles/574215/                                !!!
!!!                                                                 !!!
!!! WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING !!!
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

On Linux, SECCOMP security feature looks nice a nice start to build a Python
sandbox.

Other sandboxing projects for Python:

* PyPy project has a sandbox:
  http://pypy.org/
* http://pythonsweetness.tumblr.com/post/65442885019/secure-low-overhead-eval-sandbox-in-80-lines-of-python
* http://chdir.org/~nico/seccomp-nurse/

The old README follows.





----





pysandbox is a Python sandbox. By default, untrusted code executed in the
sandbox cannot modify the environment (write a file, use print or import a
module). But you can configure the sandbox to choose exactly which features are
allowed or not, eg. import sys module and read /etc/issue file.

Website: http://github.com/haypo/pysandbox/


Features
========

Blocked Python functionality (by default):

 * Deny access to the file system
 * Deny importing Python modules
 * Deny exiting Python
 * Deny access to stdin, stdout or stderr
 * Deny some builtin symbols like execfile(), reload() or KeyboardInterrupt
 * Deny execution of arbitrary bytecode (creation of arbitrary code object)

You can enable all of these features by setting the sandbox configuration.

By default, the untrusted code is executed in a subprocess with the following
limits:

 * timeout = 5 seconds
 * memory limit = 200 MB
 * recursion limit = 50 frames
 * number of child processes = 0 (forking and threads are disabled at the OS level)
 * pysandbox is able to catch crashes like segmentation fault (SIGSEGV)
 * stdin, stdout and stderr are redirected to /dev/null (or :NUL on Windows)
 * input and output data are limited to 64 KB

Protection of the namespace:

 * Deny access to function closure, globals, defaults and code
 * Deny access to frame locals
 * Deny access to types' subclasses
 * __builtins__ is read only
 * Deny access to dict methods able to modify a dict, eg. dict.__setitem__.
   But you can use "d[key] = value" and "del d[key]" instead
 * Use a whitelist for sys.path


Limitations
===========

pysandbox is a sandbox for the Python namespace, not a sandbox between Python
and the operating system. It does not protect your system against Python
security vulnerabilities, i.e. vulnerabilities in modules and functions
available in your sandbox (depends on your sandbox configuration). By default,
only a few functions are exposed to the sandbox namespace which limits the
attack surface.

See the Lib/test/crashers/ directory in the CPython source code to see examples
of known bugs crashing the CPython interpreter.


Configuration
=============

Use SandboxConfig class to configure your sandbox. Features are the most simple
way to configure it.

Features
--------

To enable a feature, use SandboxConfig('feature1', 'feature2', ...) or
config.enable('feature'). Available features:

 - "codecs": codecs module
 - "datetime": datetime module
 - "encodings": encodings module with ascii, latin_1, utf_8, utf_16_be,
   utf_32_be and rot_13 codecs (submodules). Enable codecs feature.
 - "exit": sys.exit(), BaseException, KeyboardInterrupt, SystemExit, quit()
 - "future": from __future__ import ...
 - "hashlib": hashlib module.
 - "help": pydoc.help(), use "import pydoc" outside the sandbox to use it. Enable regex feature.
 - "itertools": itertools module
 - "math": math module
 - "random": random module. Enable hashlib and math features.
 - "regex": compile regex, match regex, search regex, etc. (re module)
 - "site": allow to read the license file
 - "stdin": sys.stdin, input() and raw_input()
 - "stdout", "stderr": sys.stdout and sys.stderr
 - "time": time module (except sleep, strptime and tzset functions)
 - "traceback": compile() builtin, frame.f_code. Next calls to allowModule()
   will add the module filename to the open() whitelist, so Python can display
   a traceback with the source code. This feature has to be enabled before all
   other features.
 - "unicodedata": unicodedata module, required for u'\N{ATOM SYMBOL}' syntax

CPython restricted mode
-----------------------

WARNING: CPython restricted mode is unsafe because it is possible to execute
arbitrary bytecode.

Use SandboxConfig(cpython_restricted=True) to enable CPython restricted mode.
In this mode, reading a file and modifying a class are blocked. Some attributes
are hidden (eg. method.__self__), others are read only (eg. func.__doc__).

CPython restricted mode is disabled by default. The restricted mode is
incompatible with SandboxConfig's "traceback" feature and allowPath() method.

The restricted mode doesn't exist in Python3 anymore; it was removed with
the bastion and rexec modules:

 * http://svn.python.org/view?view=rev&revision=55301
 * http://hg.python.org/cpython/rev/f60c877d52c8/

Disable subprocess
------------------

It is possible to not run the untrusted code in a subprocess using
SandboxConfig(use_subprocess=False). This mode is less secure; the following
protections are disabled:

 * timeout
 * memory limit
 * number the process is not limit (forking and threads are allowed by the OS)
 * crashes aren't be caught

Other options
-------------

 - config.sys_path: trusted path list used to import modules
 - config.allowPath(path) allows reading a file from the specified path
 - config.allowModule(name, symbol1, symbol2, ...) allows importing the
   specified module, but only gives access to the specified symbols


Example
=======

With call() method: ::

    from sandbox import Sandbox

    def func(a, b):
        return a + b

    sandbox = Sandbox()
    print sandbox.call(func, 1, 2)

With execute() method: ::

    from sandbox import Sandbox, SandboxConfig
    sandbox = Sandbox(SandboxConfig('stdout'))
    sandbox.execute('print("Code executed in the sandbox")')

execute() with a local variable: ::

    from sandbox import Sandbox, SandboxConfig
    sandbox = Sandbox(SandboxConfig('stdout'))
    sandbox.execute('print(data)', locals={'data': [1, 2, 3]})    # ok
    sandbox.execute('data.append(4)', locals={'data': [1, 2, 3]}) # error

Objects passed to .call() globals/locals and .execute() arguments are
proxified: they are replaced by read-only views of the objects.


Status
======

pysanbox 1.5 is tested on Python 2.5 and 2.6 on Debian Sid.

See TODO file for the complete status.


See also
========

Python
------

 * http://wiki.python.org/moin/SandboxedPython
 * tav CPython patches:
   http://codereview.appspot.com/20051
   http://codereview.appspot.com/21043
 * secure*.py in plexnet
   http://github.com/tav/plexnet/tree/master/source/plexnet/util
 * Security in Python Wiki:
   http://wiki.python.org/moin/Security
 * safelite.py:
   http://tav.espians.com/a-challenge-to-break-python-security.html
 * Zope security:
   http://pypi.python.org/pypi/RestrictedPython
   http://svn.zope.org/zope.security/trunk/src/zope/security/
 * Brett Canon's "objcap" secured Python interpreter
   http://mail.python.org/pipermail/python-dev/2006-June/066344.html
   http://sayspy.blogspot.com/2007/05/i-have-finished-securing-python.html
   http://svn.python.org/view/python/branches/bcannon-objcap/secure_python.c?revision=56111&view=markup
 * Python taint mode:
   http://www.cats-muvva.net/software/
 * Controlling Access to Resources Within The Python Interpreter:
   http://www.cs.ubc.ca/~drifty/papers/python_security.pdf
 * PyPy sandbox:
   http://codespeak.net/pypy/dist/pypy/doc/sandbox.html
 * mxProxy:
   http://www.egenix.com/products/python/mxBase/mxProxy/
 * Python 2.3: rexec and Bastion

Python-dev mailing list
-----------------------

 * "Python jail: whitelist vs blacklist"
   Victor Stinner, Tue Feb 24 13:50:40 CET 2009
   http://mail.python.org/pipermail/python-dev/2009-February/086444.html
 * "Challenge: Please break this!"
   tav, Mon Feb 23 23:41:30 CET 2009
   http://mail.python.org/pipermail/python-dev/2009-February/086401.html
   http://mail.python.org/pipermail/python-dev/2009-February/086413.html
   http://mail.python.org/pipermail/python-dev/2009-February/086439.html
 * "Reviving restricted mode?"
   Guido van Rossum, Sun Feb 22 17:45:27 CET 2009
   http://mail.python.org/pipermail/python-dev/2009-February/086352.html
 * "object capability; func_closure; __subclasses__"
   tav, Thu Jun 28 03:04:42 CEST 2007
   http://mail.python.org/pipermail/python-dev/2007-June/073724.html
 * "Capabilities"
   Guido van Rossum, Fri, 07 Mar 2003 12:41:16 -0500
   http://mail.python.org/pipermail/python-dev/2003-March/033820.html
   http://mail.python.org/pipermail/python-dev/2003-March/033854.html
   ...
   (read the whole archive of march and april 2003)

Other
-----

 * Python shell running in Google AppEngine (which uses a sandbox)
   http://shell.appspot.com/
 * http://lua-users.org/wiki/SandBoxes
 * "Capability-based Financial Instruments"
   Mark S. Miller, Chip Morningstar and Bill Frantz, 2000
   http://www.erights.org/elib/capability/ode/index.html