How can I sandbox Python in pure Python?

This is really non-trivial.

There are two ways to sandbox Python. One is to create a restricted environment (i.e., very few globals etc.) and exec your code inside this environment. This is what Messa is suggesting. It's nice but there are lots of ways to break out of the sandbox and create trouble. There was a thread about this on Python-dev a year ago or so in which people did things from catching exceptions and poking at internal state to break out to byte code manipulation. This is the way to go if you want a complete language.

The other way is to parse the code and then use the ast module to kick out constructs you don't want (e.g. import statements, function calls etc.) and then to compile the rest. This is the way to go if you want to use Python as a config language etc.

Another way (which might not work for you since you're using GAE), is the PyPy sandbox. While I haven't used it myself, word on the intertubes is that it's the only real sandboxed Python out there.

Based on your description of the requirements (The requirements are support for variables, basic conditionals and function calls (not definitions)) , you might want to evaluate approach 2 and kick out everything else from the code. It's a little tricky but doable.

AFAIK it is possible to run a code in a completely isolated environment:

exec somePythonCode in {'__builtins__': {}}, {}

But in such environment you can do almost nothing :) (you can not even import a module; but still a malicious user can run an infinite recursion or cause running out of memory.) Probably you would want to add some modules that will be the interface to you game engine.

Roughly ten years after the original question, Python 3.8.0 comes with auditing. Can it help? Let's limit the discussion to hard-drive writing for simplicity - and see:

from sys import addaudithook
def block_mischief(event,arg):
    if 'WRITE_LOCK' in globals() and ((event=='open' and arg[1]!='r') 
            or event.split('.')[0] in ['subprocess', 'os', 'shutil', 'winreg']): raise IOError('file write forbidden')

addaudithook(block_mischief)

So far exec could easily write to disk:

exec("open('/tmp/FILE','w').write('pwned by l33t h4xx0rz')", dict(locals()))

But we can forbid it at will, so that no wicked user can access the disk from the code supplied to exec(). Pythonic modules like numpy or pickle eventually use the Python's file access, so they are banned from disk write, too. External program calls have been explicitly disabled, too.

WRITE_LOCK = True
exec("open('/tmp/FILE','w').write('pwned by l33t h4xx0rz')", dict(locals()))
exec("open('/tmp/FILE','a').write('pwned by l33t h4xx0rz')", dict(locals()))
exec("numpy.savetxt('/tmp/FILE', numpy.eye(3))", dict(locals()))
exec("import subprocess; subprocess.call('echo PWNED >> /tmp/FILE', shell=True)",     dict(locals()))

An attempt of removing the lock from within exec() seems to be futile, since the auditing hook uses a different copy of locals that is not accessible for the code ran by exec. Please prove me wrong.

exec("print('muhehehe'); del WRITE_LOCK; open('/tmp/FILE','w')", dict(locals()))

...
OSError: file write forbidden

Of course, the top-level code can enable file I/O again.

del WRITE_LOCK
exec("open('/tmp/FILE','w')", dict(locals()))

Sandboxing within Cpython has proven extremely hard and many previous attempts have failed. This approach is also not entirely secure e.g. for public web access:

1. perhaps hypothetical compiled modules that use direct OS calls cannot be audited by Cpython - whitelisting the safe pure pythonic modules is recommended.

2. Definitely there is still the possibility of crashing or overloading the Cpython interpreter.

3. Maybe there remain even some loopholes to write the files on the harddrive, too. But I could not use any of the usual sandbox-evasion tricks to write a single byte. We can say the "attack surface" of Python ecosystem reduces to rather a narrow list of events to be (dis)allowed: https://docs.python.org/3/library/audit_events.html

I would be thankful to anybody pointing me to the flaws of this approach.

EDIT: So this is not safe either! I am very thankful to @Emu for his clever hack using exception catching and introspection:

#!/usr/bin/python3.8
from sys import addaudithook
def block_mischief(event,arg):
    if 'WRITE_LOCK' in globals() and ((event=='open' and arg[1]!='r') or event.split('.')[0] in ['subprocess', 'os', 'shutil', 'winreg']):
        raise IOError('file write forbidden')

addaudithook(block_mischief)
WRITE_LOCK = True
exec("""
import sys
def r(a, b):
    try:
        raise Exception()
    except:
        del sys.exc_info()[2].tb_frame.f_back.f_globals['WRITE_LOCK']
import sys
w = type('evil',(object,),{'__ne__':r})()
sys.audit('open', None, w)
open('/tmp/FILE','w').write('pwned by l33t h4xx0rz')""", dict(locals()))

I guess that auditing+subprocessing is the way to go, but do not use it on production machines:

https://bitbucket.org/fdominec/experimental_sandbox_in_cpython38/src/master/sandbox_experiment.py

I'm not sure why nobody mentions this, but Zope 2 has a thing called Python Script, which is exactly that - restricted Python executed in a sandbox, without any access to filesystem, with access to other Zope objects controlled by Zope security machinery, with imports limited to a safe subset.

Zope in general is pretty safe, so I would imagine there are no known or obvious ways to break out of the sandbox.

I'm not sure how exactly Python Scripts are implemented, but the feature was around since like year 2000.

And here's the magic behind PythonScripts, with detailed documentation: http://pypi.python.org/pypi/RestrictedPython - it even looks like it doesn't have any dependencies on Zope, so can be used standalone.

Note that this is not for safely running arbitrary python code (most of the random scripts will fail on first import or file access), but rather for using Python for limited scripting within a Python application.

_{This answer is from my comment to a question closed as a duplicate of this one: Python from Python: restricting functionality?}