Dynamic code execution with `exec` and `eval`

Syntax#

• eval(expression[, globals=None[, locals=None]])
• exec(object)
• exec(object, globals)
• exec(object, globals, locals)

Parameters#

Remarks#

In exec, if globals is locals (i.e. they refer to the same object), the code is executed as if it is on the module level. If globals and locals are distinct objects, the code is executed as if it were in a class body.

If the globals object is passed in, but doesn’t specify __builtins__ key, then Python built-in functions and names are automatically added to the global scope. To suppress the availability of functions such as print or isinstance in the executed scope, let globals have the key __builtins__ mapped to value None. However, this is not a security feature.

The Python 2 -specific syntax shouldn’t be used; the Python 3 syntax will work in Python 2. Thus the following forms are deprecated:

Evaluating an expression with eval

>>> expression = '5 + 3 * a'
>>> a = 5
>>> result = eval(expression)
>>> result
20

Precompiling an expression to evaluate it multiple times

compile built-in function can be used to precompile an expression to a code object; this code object can then be passed to eval. This will speed up the repeated executions of the evaluated code. The 3rd parameter to compile needs to be the string 'eval'.

>>> code = compile('a * b + c', '<string>', 'eval')
>>> code
<code object <module> at 0x7f0e51a58830, file "<string>", line 1>
>>> a, b, c = 1, 2, 3
>>> eval(code)
5

Evaluating an expression with eval using custom globals

>>> variables = {'a': 6, 'b': 7}
>>> eval('a * b', globals=variables)
42

As a plus, with this the code cannot accidentally refer to the names defined outside:

>>> eval('variables')
{'a': 6, 'b': 7}
>>> eval('variables', globals=variables)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<string>", line 1, in <module>
NameError: name 'variables' is not defined

Using defaultdict allows for example having undefined variables set to zero:

>>> from collections import defaultdict
>>> variables = defaultdict(int, {'a': 42})
>>> eval('a * c', globals=variables)  # note that 'c' is not explicitly defined
0

Evaluating a string containing a Python literal with ast.literal_eval

If you have a string that contains Python literals, such as strings, floats etc, you can use ast.literal_eval to evaluate its value instead of eval. This has the added feature of allowing only certain syntax.

>>> import ast
>>> code = """(1, 2, {'foo': 'bar'})"""
>>> object = ast.literal_eval(code)
>>> object
(1, 2, {'foo': 'bar'})
>>> type(object)
<class 'tuple'>

However, this is not secure for execution of code provided by untrusted user, and it is trivial to crash an interpreter with carefully crafted input

>>> import ast
>>> ast.literal_eval('()' * 1000000)
[5]    21358 segmentation fault (core dumped)  python3

Here, the input is a string of () repeated one million times, which causes a crash in CPython parser. CPython developers do not consider bugs in parser as security issues.

Executing code provided by untrusted user using exec, eval, or ast.literal_eval

It is not possible to use eval or exec to execute code from untrusted user securely. Even ast.literal_eval is prone to crashes in the parser. It is sometimes possible to guard against malicious code execution, but it doesn’t exclude the possibility of outright crashes in the parser or the tokenizer.

To evaluate code by an untrusted user you need to turn to some third-party module, or perhaps write your own parser and your own virtual machine in Python.