Sandboxed Python

See also: How can I run an untrusted Python script safely (i.e. Sandbox) (from the Asking for Help page), Security.

Is there such a thing as "Sandboxed Python"? Something where you can run a Python that is limited in what it can do.

You could do all sorts of cool things with a sandboxed Python:

Technical Mechanisms

A "Sandboxed Python" would let you permit or forbid modules, limit execution slices, permit or deny network traffic, constrain filesystem access to a particular directory (floated as "/"), and so on. It is also referred to as RestrictedExecution, a topic brought up by Mitch Kapor at PyCon and noted on his blog.

Related topics include capabilities as mentioned in PythonThreeDotOh, specifically this message on "capability-mediated modules". Some work has been done to support capabilities in languages with similar heritage to Python, notably CaPerl.

Implementation/Realisation Strategies

Some of these strategies were suggested in the discussion below.

Use Another Runtime

PyPy has support for creating a sandboxed Python interpreter. The Java and CLR/.NET runtimes support restricted execution, and these can be utilised through the Jython and IronPython variants of Python (as well as by other languages, obviously).

Use Operating System Support

Some operating systems support various levels of sandboxing, with the most extreme form being that of virtualisation or emulation. Yet UNIX-oriented operating systems also provide more lightweight mechanisms for restricting process behaviour such as chroot "jails", mandatory access controls, resource limits, and so on.

Modify the CPython Runtime

Whilst suggestions about removing functionality from CPython have been generally rejected either as being too simplistic and naive with regard to the possibilities to subvert Python's introspection mechanisms, or too restrictive with regard to support for commonly used modules or language features, it is at least reasonable to entertain the notion that a version of CPython linked to special system libraries and built without the more "powerful" or "unsafe" extensions (see Plash's sandboxing and libc call modifications for one approach) would not have the ability to perform arbitrary operations on the filesystem or network. Indeed, this is the basis on which operating systems are hosted within each other using various virtualisation technologies.

Use a Sandbox as an Accessory

In some cases, the functionality of untrusted code may be strictly defined. For example, mobile agent systems may, for the agents themselves, permit only a highly restricted style of program, or may only permit a very limited style of interaction with other components. One solution which can be considered within this particular domain is Zope's RestrictedPython - restricted subset of Python - although that solution focuses on reducing the level of expression in the language subset employed.

History

Prior to Python 2.3, there were two modules, rexec and Bastion which were attempted to provide a sandboxed environment. Unfortunately there were known bugs and work-arounds in these modules, and the introduction of new-style classes in Python 2.3 simply made matters worse. Rather than provide a false sense of security, the modules were removed in Python 2.3.

There are a number of people who are interested in creating a new sandboxing system for Python. Apart from the Zope RestrictedPython implementation, some work by Brett Cannon is ongoing, and related works such as mxProxy are also available (although the objectives of the latter and similar projects may be less ambitious than providing a full restricted execution environment).


Reactions

For my part, I think this is something that the PSF should fund development on so that it happens sooner rather than later. There are enough interested parties, that lends itself to getting government or other grant funding. -- KevinAltis

Having used the rexec stuff years ago (1995-6), I seem to recall that the principles of that module involved restricting the modules available, possibly along with the attributes available from each module, class and object. Is there a concise summary of how such mechanisms could have been subverted/exploited and a description of why they are particularly hard to fix? Or is that what this page is about? -- PaulBoddie

I think it should be what the page is about. The basic problem is that Python's introspection allows several ways to "escape". For instance, code which is passed a file object, f, are given access to that single file, but it ALSO can use "type(f)", or "f.class" to get access to "file" itself, and then use it to open any file. Security experts will tell you that "plugging holes" is generally a poor approach to security because there usually turn out to be additional "holes" you hadn't considered -- true security requires an overall design with security considerations included. Doing this for Python is not impossible, but may prove difficult. On the other hand, I would agree that it is worthwhile. -- MichaelChermside

I also have another idea... I have long thought that much of the issues could be handled by creating a module which would make it easy to run bits of Python code in a separate interpreter, running as a separate process, and sandboxed using OS features. The difficulty would be that such a module would require OS-specific code... but it would have the potential to provide sandboxing that was as reliable as the OS would allow. -- MichaelChermside

Sounds like Parrot, (PythonAndParrot,) to me. -- LionKimbro

How so? -- MichaelChermside

Oh, sorry. No, Parrot doesn't refer to running bits of Python code in a seperate interpreter.

Parrot seems like a good target for running the sandboxed code in. I mean: ".NET" and all. Make it so you can secure the Parrot interpreter, and you're good. If a module requires custom C code links, or whatever, make it so you can close the gates to individual parts of C code, or something. And then make it so you can put various RestrictedExecution thingies on the Python code. -- LionKimbro

Perhaps a better solution with the same approach would be Jython. Java provides excellent sandboxing capabilities, and Jython has the distinct advantage that it already exists and works. -- MichaelChermside

Whereas Parrot; I think Parrot will function just fine. It'll just take a little time. -- LionKimbro 2004-05-07 16:54:04

We created a project Pynbox to run Python in a NativeClient OS-level sandbox. It makes it easy to install (root not required), works cross-platform, supports native modules, and allows configuring read/write access to specific filesystem directories. We use it at Grist Labs, but would love to know if there is wider interest and to get feedback. -- DmitryS 2017-06-06

SandboxedPython (last edited 2017-06-06 20:21:57 by DmitryS)

Unable to edit the page? See the FrontPage for instructions.