Cisco launches the open-source Model Provenance Kit to strengthen AI supply-chain security by helping organizations verify the origin and integrity of AI models, framing provenance as a foundational layer of AI governance.

The kit aims to evolve into a de facto standard for model traceability through broader industry participation, dataset expansion, and open collaboration.

Early results show related model pairs cluster around a provenance score near 1.0, while unrelated pairs stay below 0.70, signaling reliable discrimination between related and unrelated models.

The initial fingerprint database contains about 150 base models from more than 45 families and 20 publishers, covering a wide parameter range to support scan mode.

The kit uses two-stage analysis: a fast metadata-based architectural screen followed by a weight-level analysis across five signals to assess provenance.

Provenance is treated like a DNA test, comparing both metadata and learned parameters to verify common origin and detect modifications.

A composite fingerprint—combining tokenizer similarity, embedding geometry, normalization-layer traits, energy profiles, and weight comparisons—helps trace origins and relationships and is harder to spoof than metadata alone.

The toolkit traces lineage through metadata, architecture, and learned parameters to identify modifications, fine-tuning from base models, or potential compromises.

The initiative comes as Hugging Face hosts over 2 million public models and more than 13 million users, underscoring the challenge of distinguishing high-quality models from compromised ones.

Practical details: the pipeline runs on CPU, architectural matches resolve quickly, and the repository and fingerprint dataset are available on GitHub and Hugging Face.

The toolkit addresses risks such as model tampering, poisoned datasets, regulatory and licensing gaps, and broader AI supply-chain integrity concerns by enabling provenance tracking.

It is built as a Python-based CLI tool with a growing fingerprint dataset hosted on Hugging Face.