Avoid dangerous file parsing and object serialization libraries

Many common libraries that are often used for reading configuration files and deserializing objects are very dangerous because they can allow execution of arbitrary code. By default, libraries such as PyYAML and pickle do not provide strong separation of data and code, and thus allow code to be embedded inside the input.

Often the input to these libraries is untrusted or only partially trusted. These unsafe inputs can come from configuration files or be provided via REST APIs. For example, we often use YAML for configuration files but YAML files can also contain embedded Python code. This may provide an attacker with a method to execute code.

Many, but not all, of these libraries, offer safe interfaces that disable features that enable code execution. You always want to use the safe functions to load input. Often the obvious function to use is not the safe one and we should check the documentation for libraries not covered here.

Python Libraries

We often use YAML, pickle, or eval to load data into our Python programs, but this is dangerous. PyYAML has a safe way to load code, but pickle and eval do not.

Module

Problem

Use

Avoid

PyYAML

Allows creating arbitrary Python objects.

yaml.safe_load

yaml.load

pickle

Allows creating arbitrary Python objects.

Do not use

pickle.load, pickle.loads

cPickle

Allows creating arbitrary Python objects.

Do not use

cPickle.load, cPickle.loads

eval

Runs all input as Python code

Do not use

eval

exec

Runs all input as Python code (Python 3.x)

Do not use

exec

Incorrect

yaml.load is the obvious function to use but it is dangerous:

import yaml
import pickle
conf_str = '''
!!python/object:__main__.AttackerObj
key: 'value'
'''
conf = yaml.load(conf_str)

Using pickle or cPickle with untrusted input can result in arbitrary code execution.

import pickle
import cPickle

user_input = "cos\nsystem\n(S'cat /etc/passwd'\ntR.'\ntR."
cPickle.loads(user_input) # results in code execution
pickle.loads(user_input)  # results in code execution

Similarly eval and exec are difficult to use safely with input that comes from an untrusted source.

user_input = "os.system('cat /etc/passwd')"
eval(user_input) # execute python expressions

user_input = "import os; os.system('cat /etc/passwd')"
exec(user_input) # execute _any_ python code

Correct

Here we use PyYAMLs safe YAML loading function:

import yaml
conf_str = '''
- key: 'value'
- key: 'value'
'''
conf = yaml.safe_load(conf_str)

There is no safe alternative for pickle.load. However in most cases using pickle for serialization of data objects is something that can be avoided altogether.

Consequences

  • Anyone that can control the input passed to dangerous libraries can gain arbitrary code execution on the system running the dangerous library.

References