asteval Sandbox Escape: arbitrary native memory read/write via numpy ctypes in default asteval Interpreter

Summary

With its default configuration (numpy enabled, import disabled), asteval's Interpreter lets an attacker-controlled expression obtain a raw arbitrary process-memory read and write primitive, without using import, any __dunder__ attribute, or eval/exec/getattr. Arbitrary in-process read/write is equivalent to arbitrary code execution and is a complete escape of the sandbox whose entire purpose is "untrusted string in, no arbitrary execution out." Any application that feeds untrusted input to asteval with numpy installed (the default) is affected.

Details

asteval's attribute filter (asteval/astutils.py: safe_getattr) blocks every __dunder__ name and blocks objects whose attribute value is identity-equal to one of the modules in UNSAFE_MODULES = {io, os, sys, ctypes}. The ctypes module entry was added recently (commit 9d9d430) and correctly blocks ndarray.ctypes._ctypes.

However, the module check is identity-only against the ctypes module. It does not cover ctypes type objects and their metaclass methods, which are reachable through numpy's ndarray.ctypes wrapper using only ordinary (non-dunder) attribute names:

zeros(1, dtype=int32).ctypes.shape._type_      ->  <class 'ctypes.c_long'>

ndarray.ctypes exposes .shape (a ctypes array) whose element type ._type_ is ctypes.c_long. None of ctypes, .shape, ._type_ is a dunder, none is in UNSAFE_ATTRS, and the returned value is a type, not the ctypes module, so safe_getattr permits all of them.

On that ctypes type, the metaclass method from_address is reachable (non-dunder, not in UNSAFE_ATTRS; it is not even listed by dir(), which is likely why it was missed):

• Arbitrary read: c_long.from_address(addr).value reads 8 bytes at any address. id() (a permitted builtin) supplies arbitrary object addresses.
• Arbitrary write: cell = c_long.from_address(addr); cell.value = X writes 8 bytes to any address. The write half rides asteval's unfiltered setattr in Interpreter.node_assign (the ast.Attribute branch performs setattr(self.run(node.value), node.attr, val) with no attribute-name check).

Root cause is two gaps:

1. safe_getattr blocks the ctypes module but not ctypes types / metaclass methods (from_address, from_buffer, from_buffer_copy, in_dll, from_param) reachable via ndarray.ctypes ... ._type_.
2. node_assign performs attribute writes (setattr) and deletes (delattr) with no attribute-name filtering.

This belongs to the known "numpy is a large attack surface" class (the docs already note open() read and ndarray.tofile() write), but this specific arbitrary memory read/write chain is undocumented and bypasses the most recent ctypes-module hardening. All previously reported escapes (CVE-2025-24359 / GHSA-3wwr-3g9f-9gc7, GHSA-vp47-9734-prjw, reduce/reduce_ex, classic __subclasses__ traversal) are patched on the current code; this one is live.

PoC

Self contained POC here: https://gist.github.com/thegr1ffyn/16b67c5f9b5339a7e2bdc91423ff09e3 Environment: pip install asteval numpy (verified on asteval 1.0.8, numpy 2.4.6, CPython 3.12.3; the chain is numpy-1.x/2.x robust). Default Interpreter (use_numpy=True, import disabled).

Minimal one-expression arbitrary read (reads 8 bytes at an attacker-chosen address):

zeros(1,dtype=int32).ctypes.shape._type_.from_address(id(zeros(1))).value

Minimal arbitrary write (writes 0x4142434445464748 to a chosen address; here our own array buffer, observed back through numpy):

a = zeros(2, dtype=int32)
cell = a.ctypes.shape._type_.from_address(a.ctypes.data)
cell.value = 0x4142434445464748        # -> a[0]=0x45464748, a[1]=0x41424344

A full self-contained script is attached (poc_asteval_ctypes.py); running it prints the recovered PyObject header of a private object (arbitrary read) and confirms a raw write landing at a chosen pointer (arbitrary write), all from a default, import-disabled interpreter.

Impact

Sandbox escape / protection-mechanism failure leading to arbitrary in-process native memory read and write (RCE-equivalent). Impact:

• Disclosure of any data in the host process's address space (secrets, keys, other users' data).
• Corruption of arbitrary memory -> control-flow hijack / arbitrary code execution and/or process crash (DoS).

Affected: any application that evaluates untrusted/attacker-influenced expressions with asteval while numpy is installed (the default). No authentication and no special configuration is required; import does not need to be enabled. Mitigation until patched: construct the interpreter with use_numpy=False.