Back to Gap Map Product / UX / Safety & Guardrails

ModelScan

Protect AI
open source / Overall score: 2.5

ModelScan checks machine-learning model files for code that would run when the model is loaded, an attack known as model serialization. It reads Pickle, H5, Keras and SavedModel files without loading them, covering models from PyTorch, TensorFlow, Keras, scikit-learn and XGBoost, and ranks what it finds by severity. Protect AI publishes it as a command-line tool and Python package.

Openness

5 medium confidence
5.0
license
Apache-2.0 (OSI)
source
public (the full scanner and CLI)
core features withheld
no — ModelScan is a complete scanner that runs locally; Protect AI's Guardian is a separate enterprise product, not a withheld ModelScan module

ModelScan's code is Apache-2.0, and the published scanner is complete and runs locally for the formats it supports. Protect AI sells Guardian as a separate enterprise product with newer scanners and broader model support; it is not a part of ModelScan kept back for paying customers.

Adoption

2 high confidence
2.0

PyPI downloads of the modelscan package, the tool's only distribution. Scans in automated pipelines can install it again on every run, so a download is not one user.

Capability

3 medium confidence
3.0

ModelScan catches one class of attack, malicious code hidden in a model file, before the model is loaded, across the main serialization formats. It does not probe a running model for jailbreaks, leakage or toxic output, which garak does across many model providers, so it sits a step below garak.

  • https://raw.githubusercontent.com/protectai/modelscan/HEAD/README.md recorded 2026-11-12

    The README's support table marks Pickle, Protocol Buffer (SavedModel), HD5 and Keras V3 formats as supported for PyTorch, TensorFlow, Keras and classic ML libraries "pickle.dump(), dill.dump(), joblib.dump(), cloudpickle.dump()", and ranks unsafe code as CRITICAL, HIGH, MEDIUM and LOW.

Verified 2026-11-12