CodeJail

CodeJail manages execution of untrusted code in secure sandboxes. It is designed primarily for Python execution, but can be used for other languages as well.

Security is enforced with AppArmor. If your operating system doesn't support AppArmor, or if the AppArmor profile is not defined and configured correctly, then CodeJail will not protect the execution.

CodeJail is designed to be configurable, and will auto-configure itself for Python execution if you install it properly.

A CodeJail sandbox consists of several pieces:

1. Sandbox environment. For a Python setup, this would be Python and associated core packages as a virtualenv. This is denoted throughout this document as <SANDENV>. This is read-only, and shared across sandbox instantiations.

   Sandboxed code also has access to OS libraries to the extent that the AppArmor profile permits it.

2. Sandbox execution directory. This is an ephemeral read-only directory named like /tmp/codejail-XXXXXXXX containing the submitted code (./jailed_code), optional additional files, and a writable temporary directory (./tmp) that the submitted code can use as a scratch space.

   The submitted code is typically the code submitted by the student to be tested on the server, and the additional files are typically a python_lib.zip containing grading or utility libraries.

To run, CodeJail requires two user accounts. One account is the main account under which the code runs, which has access to create sandboxes. This will be referred to as <SANDBOX_CALLER>. The second account is the account under which the sandbox runs. This is typically the account sandbox.

Supported Versions

This library currently is tested to work with the following versions

Python:

• 3.11

Ubuntu:

• 22.04
• 24.04

(Note that the Python version used inside the sandbox may be different from the version used for the library itself.)

Installation

These instructions detail how to configure your operating system so that CodeJail can execute Python code safely. However, it is also possible to set codejail.safe_exec.ALWAYS_BE_UNSAFE = True and execute submitted Python directly on the machine, with no security whatsoever. This may be fine for developers' machines who are unconcerned with security, and allows testing an integration with CodeJail's API. It must not be used if any input is coming from untrusted sources, however. Do not use this option in production systems.

To secure Python execution, you'll be creating a new virtualenv. This means you'll have two: the main virtualenv for your project, and the new one for sandboxed Python code.

Choose a place for the new virtualenv, call it <SANDENV>. It will be automatically detected and used if you put it right alongside your existing virtualenv, but with -sandbox appended. So if your existing virtualenv is in /home/chris/ve/myproj, make <SANDENV> be /home/chris/ve/myproj-sandbox.

The user running the LMS is <SANDBOX_CALLER>, for example, you on your dev machine, or www-data on a server.

Other details here that depend on your configuration:

1. Create the new virtualenv, using --copies so that there's a distinct Python executable to limit:

   $ sudo python3.11 -m venv --copies <SANDENV>
   

   By default, the virtualenv would just symlink against the system Python, and apparmor's default configuration on some operating systems may prevent confinement from being appled to that.

2. (Optional) If you have particular packages you want available to your sandboxed code, install them by activating the sandbox virtual env, and using pip to install them:

   $ <SANDENV>/bin/pip install -r requirements/sandbox.txt
   

3. Add a sandbox user:

   $ sudo addgroup sandbox
   $ sudo adduser --disabled-login sandbox --ingroup sandbox
   

4. Let the web server run the sandboxed Python as sandbox. Create the file /etc/sudoers.d/01-sandbox:

   $ sudo visudo -f /etc/sudoers.d/01-sandbox
   
   <SANDBOX_CALLER> ALL=(sandbox) SETENV:NOPASSWD:<SANDENV>/bin/python
   <SANDBOX_CALLER> ALL=(sandbox) SETENV:NOPASSWD:/usr/bin/find
   <SANDBOX_CALLER> ALL=(ALL) NOPASSWD:/usr/bin/pkill
   

   (Note that the find binary can run arbitrary code, so this is not a safe sudoers file for non-codejail purposes.)

5. Edit an AppArmor profile. This is a text file specifying the limits on the sandboxed Python executable. The file must be in /etc/apparmor.d and should be named based on the executable, with slashes replaced by dots. For example, if your sandboxed Python is at /home/chris/ve/myproj-sandbox/bin/python, then your AppArmor profile must be /etc/apparmor.d/home.chris.ve.myproj-sandbox.bin.python.

   See sample profile in apparmor-profiles/. The profile must be customized to match your sandbox location.

6. Parse the profiles:

   $ sudo apparmor_parser --replace --warn=all --warn=no-debug-cache --Werror <APPARMOR_FILE>
   

7. Reactivate your project's main virtualenv again.

8. Disable using PAM to set rlimits:

   sed -i '/pam_limits.so/d' /etc/pam.d/sudo
   

Using CodeJail

If your CodeJail is properly configured to use safe_exec, try these commands at your Python terminal:

import codejail.jail_code
codejail.jail_code.configure('python', '<SANDENV>/bin/python', user='sandbox')
import codejail.safe_exec
jailed_globals = {}
codejail.safe_exec.safe_exec("output=open('/etc/passwd').read()", jailed_globals)
print(jailed_globals)  # should be unreachable if codejail is working properly

This should fail with an exception.

If you need to change the packages installed into your sandbox's virtualenv, you'll need to disable AppArmor, because your sandboxed Python doesn't have the rights to modify the files in its site-packages directory.

1. Disable AppArmor for your sandbox:

   $ sudo apt-get install apparmor-utils  # if you haven't already
   $ sudo aa-complain /etc/apparmor.d/home.chris.ve.myproj-sandbox.bin.python
   

2. Install or otherwise change the packages installed:

   $ pip install -r requirements/sandbox.txt
   

3. Re-enable AppArmor for your sandbox:

   $ sudo aa-enforce /etc/apparmor.d/home.chris.ve.myproj-sandbox.bin.python
   

Tests

To run tests, you must perform the standard installation steps. Then you must set the following environment variables:

$ export CODEJAIL_TEST_USER=<owner of sandbox (usually 'sandbox')>
$ export CODEJAIL_TEST_VENV=<SANDENV>

Run the tests with the Makefile:

$ make tests

Several proxy tests are skipped if proxy mode is not configured.

Design

CodeJail is general-purpose enough that it can be used in a variety of projects to run untrusted code. It provides two layers:

• jail_code.py offers secure execution of subprocesses. It does this by running the program in a subprocess managed by AppArmor.
• safe_exec.py offers specialized handling of Python execution, using jail_code to provide the semantics of Python's exec statement.

CodeJail runs programs under AppArmor. AppArmor is an OS-provided feature to limit the resources programs can access. To run Python code with limited access to resources, we make a new virtualenv, then name that Python executable in an AppArmor profile, and restrict resources in that profile. CodeJail will execute the provided Python program with that executable, and AppArmor will automatically limit the resources it can access. CodeJail also uses setrlimit to limit the amount of CPU time and/or memory available to the process.

codejail.jail_code takes a program to run, files to copy into its environment, command-line arguments, and a stdin stream. It creates a temporary directory, creates or copies the needed files, spawns a subprocess to run the code, and returns the output and exit status of the process.

codejail.safe_exec emulates Python's exec statement. It takes a chunk of Python code, and runs it using jail_code, modifying the globals dictionary as a side-effect. safe_exec does this by serializing the globals into and out of the subprocess as JSON.

Limitations

• If codejail or AppArmor is not configured properly, codejail may default to running code insecurely (no sandboxing). It is not secure by default. Projects integrating codejail should consider including a runtime test suite that checks for proper confinement at startup before untrusted inputs are accepted.
• Sandbox isolation is achieved via AppArmor confinement. Codejail facilitates this, but cannot isolate execution without the use of AppArmor.
• Resource limits can only be constrained using the mechanisms that Linux's rlimit makes available. Some notable deficiencies:
  • While rlimit's FSIZE can limit the size of any one file that a process can create, and can limit the number of files it has open at any one time, it cannot limit the total number of files written, and therefore cannot limit the total number of bytes written across all files. A partial mitigation is to constrain the max execution time. (All files written in the sandbox will be deleted at end of execution, in any case.)
  • The NPROC limit constrains the ability of the current process to create new threads and processes, but the usage count (how many processes already exist) is the sum across all processes with the same UID, even in other containers on the same host where the UID may be mapped to a different username. This constraint also applies to the app user due to how the rlimits are applied. Even if a UIDs are chosen so they aren't used by other software on the host, multiple codejail sandbox processes on the same host will share this usage pool and can reduce each other's ability to create processes. In this situation, NPROC will need to be set higher than it would be for a single codejail instance taking a single request at a time.
• Sandboxes do not have strong isolation from each other. Under proper configuration, untrusted code should not be able to discover other actively running code executions, but if this assumption is violated then one sandbox could theoretically interfere with another one.

Reporting Security Issues

Please do not report security issues in public. Please email security@openedx.org.