classification
Title: ast.literal_eval supports non-literals in Python 3
Type: enhancement Stage: patch review
Components: Documentation Versions: Python 3.7
process
Status: open Resolution:
Dependencies: Superseder:
Assigned To: docs@python Nosy List: David Bieber, benjamin.peterson, brett.cannon, docs@python, nascheme, ncoghlan, r.david.murray, rhettinger, serhiy.storchaka, terry.reedy
Priority: normal Keywords: patch

Created on 2017-10-12 22:31 by David Bieber, last changed 2017-11-09 21:45 by nascheme.

Pull Requests
URL Status Linked Edit
PR 4035 open serhiy.storchaka, 2017-10-18 14:34
Messages (10)
msg304294 - (view) Author: David Bieber (David Bieber) Date: 2017-10-12 22:31
# Overview
ast.literal_eval supports some non-literals in Python 3.
The behavior of this function runs contrary to the documented behavior.


# The Issue
The [documentation](https://docs.python.org/3/library/ast.html#ast.literal_eval) says of the function "It is not capable of evaluating arbitrarily complex expressions, for example involving operators or indexing."

However, literal_eval is capable of evaluating expressions with certain operators, particular the operators "+" and "-".

As has been explained previously, the reason for this is to support "complex" literals such as 2+3j. However, this has unintended consequences which I believe to be indicative of a bug. Examples of the unintended consequences include `ast.literal_eval('1+1') == 2` `ast.literal_eval('2017-10-10') == 1997`. I would expect each of these calls to literal_eval to throw a ValueError, as the input string is not a literal. Instead, literal_eval successfully evaluates the input string, in the latter case giving an unexpected result (since the intent of the string is to represent a 21st century date.)

Since issue arose as a [Python Fire issue](https://github.com/google/python-fire/issues/97), where the behavior of Python Fire was unexpected for inputs such as those described above (1+1 and 2017-10-10) only in Python 3. For context, Python Fire is a CLI library which uses literal_eval as part of its command line argument parsing procedure.

I think the resolution to this issue is having literal_eval raise a ValueError if the ast of the input represents anything other than a Python literal, as described in the documentation. That is, "The string or node provided may only consist of the following Python literal structures: strings, bytes, numbers, tuples, lists, dicts, sets, booleans, and None." Additional operations, such as the binary operations "+" and "-", unless they explicitly create a complex number, should produce a ValueError.

If that resolution is not the direction we take, I also would appreciate knowing if there is another built in approach for determining if a string or ast node represents a literal.


# Reproducing
The following code snippets produce different behaviors in Python 2 and Python 3.
```python
import ast
ast.literal_eval('1+1')
```

```python
import ast
ast.literal_eval('2017-10-10')
```


# References
- The Python Fire issue is here: https://github.com/google/python-fire/issues/97
- Prior discussion of a similar issue: https://bugs.python.org/issue22525
- I think is where complex number support was originally added: https://bugs.python.org/issue4907
- In this thread, https://bugs.python.org/issue24146, one commenter explains that literal_eval's support for "2+3" is an unintentional side effect of supporting complex literals.
msg304367 - (view) Author: Terry J. Reedy (terry.reedy) * (Python committer) Date: 2017-10-13 22:03
It has been some time since literal_eval literally only evaluated literals.  'constant_eval' might be a better name now, with the proviso of 'safely, in reasonable time'.

>>> from ast import literal_eval as le
>>> le('(1,2,3)')
(1, 2, 3)
>>> le('(1,2, (3,4))')
(1, 2, (3, 4))

I believe there was once a time when a simple tuple would be evaluated, while a nested one would not be.

"It is not capable of evaluating arbitrarily complex expressions, for example involving operators or indexing."  I do not read this as prohibiting all operators, but rather that now all will be accepted.

>>> le(2**2)
...
ValueError: malformed node or string: 4

Exponentiation of ints can take exponential time and can be used for denial of service attacks.

>>> le('2017-10-10')
1997

This is correct.  For '2017-10-10' to be a string representing a date, it must be quoted as a string in the code.

>>> le("'2017-10-10'")
'2017-10-10'

Rolling back previous enhancements would break existing code, so a deprecation period would be required.  But I would be inclined to instead update the doc to match the updated code better.  Lets see what others think.
msg304374 - (view) Author: Nick Coghlan (ncoghlan) * (Python committer) Date: 2017-10-14 05:05
I'm marking this as documentation issue for now, as the operators that literal_eval allows are solely those where constant folding support is needed to correctly handle complex and negative numbers (as noted in the original post):

```
>>> dis.dis("+1")
  1           0 LOAD_CONST               1 (1)
              2 RETURN_VALUE
>>> dis.dis("-1")
  1           0 LOAD_CONST               1 (-1)
              2 RETURN_VALUE
>>> dis.dis("1+1")
  1           0 LOAD_CONST               1 (2)
              2 RETURN_VALUE
>>> dis.dis("1+1j")
  1           0 LOAD_CONST               2 ((1+1j))
              2 RETURN_VALUE
>>> dis.dis("2017-10-10")
  1           0 LOAD_CONST               3 (1997)
              2 RETURN_VALUE
```

So the key promise that literal_eval makes is that it will not permit arbitrary code execution, but the docs should make it clearer that it *does* permit constant folding for addition and subtraction in order to handle the full range of numeric literals.

If folks want to ensure that the input string *doesn't* include a binary operator, then that currently needs to be checked separately with ast.parse:

```
>>> type(ast.parse("2+3").body[0].value) is ast.BinOp
True
>>> type(ast.parse("2017-10-10").body[0].value) is ast.BinOp
True
```

For 3.7+, that check could potentially be encapsulated as an "allow_folding=True" keyword-only parameter (where the default gives the current behaviour, while "allow_folding=False" disables processing of UnaryOp and BinOp), but the docs update is the more immediate need.
msg304379 - (view) Author: David Bieber (David Bieber) Date: 2017-10-14 06:08
# Replies

> Rolling back previous enhancements would break existing code.

I sympathize completely with the need to maintain backward compatibility. And if this is the reason that this issue gets treated only as a documentation issue, rather than as a behavior issue, I can appreciate that.

If that is the case and literal_eval is not going to only evaluate literals, then for my use case I'll need a way to determine from a string whether it represents a literal. I can implement this myself using ast.parse and walking the resulting tree, looking for non-literal AST nodes. Would such an "is_literal" function would be more appropriate in the ast module than as a one-off function in Python Fire?

> The key promise that literal_eval makes is that it will not permit arbitrary code execution.

I disagree that this is the only key promise, and here's my reasoning. The docstring has two sentences, and each one makes a promise:
1. "Safely evaluate an expression node or a string containing a Python expression."
2. "The string or node provided may only consist of the following Python literal structures: strings, bytes, numbers, tuples, lists, dicts, sets, booleans, and None."
(1) says that evaluation is safe -- this is the key promise that you reference. (2) is also a promise though, that only certain types are allowed. While one could argue that the behavior of the function is not specified for inputs violating that criteria, I think the clear correct thing to do is to raise a ValueError if the value doesn't meet the criteria. This is what was done in Python 2, where the docstring for literal_eval is these same two sentences (modulo the inclusion of bytes and sets). It's my opinion that Python 2's behavior better matches the docstring as well as the behavior implied by the function's name.


# Additional observations
1. Python 2 _does_ support parsing complex literals, but does not support treating e.g. '1+1' as a literal.
ast.literal_eval('1+1j') # works in both Python 2 and Python 3
ast.literal_eval('1j+1') # raises a ValueError in Python 2, returns 1+1j in Python 3
ast.literal_eval('1+1') # raises a ValueError in Python 2, returns 2 in Python 3

2. Python 3 supports parsing addition and subtraction at any level of nesting.
ast.literal_eval('(1, (0, 1+1+1))') # raises a ValueError in Python 2, returns (1, (0, 3)) in Python 3

In my opinion, Python 2's behavior is correct in these situations since it matches the documentation and only evals literals as defined in the documentation.

# Source
The relevant code in Python 2.7.3 is [here](https://github.com/enthought/Python-2.7.3/blob/69fe0ffd2d85b4002cacae1f28ef2eb0f25e16ae/Lib/ast.py#L67). It explicitly allows NUM +/- COMPLEX, but not even COMPLEX +/- NUM. The corresponding code for Python 3 is [here](https://github.com/python/cpython/blob/ef611c96eab0ab667ebb43fdf429b319f6d99890/Lib/ast.py#L76). You'll notice it supports adding and subtracting arbitrary numeric types (int, float, complex).


---

Thanks for your replies and for hearing me out on this issue.
msg304400 - (view) Author: Serhiy Storchaka (serhiy.storchaka) * (Python committer) Date: 2017-10-14 19:43
The support of parsing addition and subtraction at any level of nesting was added by bc95973b51abadc84960e7836ce313f12cf515cf. The commit message and NEWS entry don't contain an issue number, thus the rationale of this change is not known. Raymond, could you please explain?
msg304573 - (view) Author: R. David Murray (r.david.murray) * (Python committer) Date: 2017-10-18 11:54
"Safely evaluate an expression node or a string containing a Python expression."

The behavior you are citing matches that documentation, as far as I can see.  1+1 is an expression involving supported literals.
msg304574 - (view) Author: Serhiy Storchaka (serhiy.storchaka) * (Python committer) Date: 2017-10-18 12:01
"""
The string or node provided may only consist of the following Python literal structures: strings, bytes, numbers, tuples, lists, dicts, sets, booleans, and None.
"""

1+1 is not a literal number.

"""
It is not capable of evaluating arbitrarily complex expressions, for example involving operators or indexing.
"""
msg304580 - (view) Author: Serhiy Storchaka (serhiy.storchaka) * (Python committer) Date: 2017-10-18 14:35
PR 4035 makes ast.literal_eval() more strict.
msg305973 - (view) Author: Serhiy Storchaka (serhiy.storchaka) * (Python committer) Date: 2017-11-09 17:42
Ping.
msg305987 - (view) Author: Neil Schemenauer (nascheme) * (Python committer) Date: 2017-11-09 21:45
Just a comment on what I guess is the intended use of literal_eval(), i.e. taking a potentially untrusted string and turning it into a Python object.  Exposing the whole of the Python parser to potential attackers would make me very worried.  Parsing code for all of Python syntax is just going to be very complicated and there can easily be bugs there.  Generating an AST and then walking over it to see if it is safe is also scary.  The "attack surface" is too large.  This is similar to the Shellshock bug. If you can trust the supplier of the string then okay but I would guess that literal_eval() is going to get used for untrusted inputs.

It would be really nice to have something like ast.literal_eval() that could be used for untrusted strings.  I would implement it by writing a retricted parser.  Keep it extremely simple.  Validate it by heavy code reviews and extensive testing (e.g. fuzzing).
History
Date User Action Args
2017-11-09 21:45:10naschemesetnosy: + nascheme
messages: + msg305987
2017-11-09 17:42:57serhiy.storchakasetmessages: + msg305973
2017-10-18 17:18:46yselivanovsetnosy: - yselivanov
2017-10-18 14:35:31serhiy.storchakasetmessages: + msg304580
versions: - Python 3.6, Python 3.8
2017-10-18 14:34:06serhiy.storchakasetkeywords: + patch
stage: patch review
pull_requests: + pull_request4009
2017-10-18 12:01:29serhiy.storchakasetmessages: + msg304574
2017-10-18 11:54:54r.david.murraysetnosy: + r.david.murray
messages: + msg304573
2017-10-14 19:43:08serhiy.storchakasetnosy: + serhiy.storchaka, rhettinger
messages: + msg304400
2017-10-14 06:08:20David Biebersetmessages: + msg304379
2017-10-14 05:05:42ncoghlansetversions: + Python 3.6
nosy: + docs@python

messages: + msg304374

assignee: docs@python
components: + Documentation
2017-10-13 22:03:44terry.reedysetversions: - Python 3.4, Python 3.5, Python 3.6
nosy: + terry.reedy, brett.cannon, ncoghlan, yselivanov, benjamin.peterson

messages: + msg304367

type: behavior -> enhancement
2017-10-12 22:32:49David Biebersettype: behavior
2017-10-12 22:31:18David Biebercreate