Executing untrusted scripts in a sandboxed environment

[Chris Angelico]

On Sat, Oct 6, 2012 at 8:22 AM, Robin Krahl <me at robin-krahl.de> wrote:
> Hi all,
>
> I need to execute untrusted scripts in my Python application. To avoid security issues, I want to use a sandboxed environment. This means that the script authors have no access to the file system. They may only access objects, modules and classes that are "flagged" or "approved" for scripting.
>
> I read that I will not be able to do this with Python scripts. (See SandboxedPython page in the Python wiki [0] and several SE.com questions, e. g. [1].) So my question is: What is the best way to "embed" a script engine in a sandboxed environment that has access to the Python modules and classes that I provide?

With extreme difficulty. A while back (couple years maybe? I don't
remember), I ignored everyone's warnings and tried to make a sandboxed
Python, embedded in a C++ application. It failed in sandboxing. With
just some trivial tinkering using Python's introspection facilities, a
couple of python-list people managed to read and write files, and
other equally dangerous actions. Shortly thereafter, we solved the
problem completely... by switching to JavaScript.

Embedding CPython in an application simply doesn't afford sandboxing.
To what extent do you actually need to run untrusted Python? Can you,
for instance, sandbox the entire process (which wasn't an option for
what we were doing)? Perhaps chrooting the Python interpreter will do
what you need. But there may still be leaks, I don't know.

ChrisA