Cisco open sources Model Provenance Kit to trace AI model lineage and detect unauthorized modifications

Cisco released Model Provenance Kit as an open source Python toolkit and command-line interface that determines whether two AI models share a common origin by analyzing architecture metadata, tokenizer structure, and learned weights.

The toolkit generates a fingerprint for each model using metadata signals, tokenizer similarity, and weight-level identity signals such as embedding geometry, normalization layers, energy profiles, and direct weight comparisons, then compares them to produce a provenance score.

Without model provenance insight, organizations risk deploying poisoned or vulnerable models, face licensing and regulatory compliance gaps (including requirements from the EU AI Act and NIST's AI Risk Management Framework), and struggle to trace security incidents back to their root cause.