Deneimaiegw/modelscan-npy-bypass-poc

ModelScan NumPy Scanner Bypass PoC

WARNING

These files contain proof-of-concept pickle payloads. Do NOT load in production environments.

Vulnerability

ModelScan's scan_numpy function (in tools/picklescanner.py) uses internal NumPy APIs (_check_version, _read_array_header) that were removed in NumPy >=2.3.5. The scanner crashes with AttributeError and reports "No issues found" while the malicious file remains fully executable.

Files

• —malicious_model.npy - Malicious .npy file with pickle RCE payload
• —create_malicious_npy.py - Creates the malicious file
• —simple_poc.py - Demonstrates the bypass

Reproduction

pip install modelscan numpy>=2.3.5
python3 simple_poc.py
modelscan -p malicious_model.npy        # Reports "No issues found"
python3 -c "import numpy as np; np.load('malicious_model.npy', allow_pickle=True)"  # Executes payload
cat /tmp/numpy_bypass_pwned.txt         # Proof of execution

Affected

• —ModelScan 0.8.8 with NumPy >=2.3.5