modelscan-numpy2-failopen-npy-poc

PoC — modelscan fails open on .npy under numpy >= 2

Proof-of-concept for a huntr Model File Vulnerabilities (MFV) report against modelscan 0.8.8.

Payload is deliberately harmless: it runs echo NPY_EXEC_OK. No network access, no deletion, nothing written outside the working directory.

The issue

modelscan/tools/picklescanner.py:234 calls a private numpy API:

np.lib.format._check_version(version)   # type: ignore[attr-defined]

numpy 2.0 removed _check_version. Under numpy >= 2 the call raises AttributeError, the .npy scanner aborts — and modelscan still prints an affirmative clean verdict. The error is relegated to a secondary Errors section and the skipped count is hidden behind --show-skipped.

pip install modelscan today resolves numpy 2.x by default, so a default install returns a clean verdict for any malicious .npy.

Measured differential — identical file, only numpy differs

Reproduce

python -m venv v2 && ./v2/bin/pip install modelscan==0.8.8 numpy==2.5.1
./v2/bin/python generate_poc.py
./v2/bin/modelscan -p poc_object_array.npy      # "No issues found!"  <-- fails open

python -m venv v1 && ./v1/bin/pip install modelscan==0.8.8 numpy==1.26.4
./v1/bin/modelscan -p poc_object_array.npy      # CRITICAL: unsafe operator 'system'

Suggested fix

• Stop calling the private np.lib.format._check_version; gate on numpy.__version__ or drop the check (_read_array_header already validates the version).
• Fail closed: a scanner that raised must not yield an affirmative "No issues found" summary. Surface aborted scans in the main verdict.

Attribution

Cyfra Tech Solutions (Roman Arce Bran), Costa Rica.