MLflow's mlflow.statsmodels flavor loaded pickled model artifacts via smio.load_pickle() in _load_model() without consulting the MLFLOW_ALLOW_PICKLE_DESERIALIZATION safety control, so setting that variable to false did not prevent deserialization. A crafted model artifact therefore achieves remote code execution on load even when the operator had explicitly disabled pickle deserialization. Fixed in 3.15.0 by adding the guard to the statsmodels flavor.