Cisco released the Model Provenance Kit on April 30, an open-source tool that generates unique fingerprints for AI models to trace their origins and flag inherited vulnerabilities. The toolkit compares models against a database of approximately 150 foundational models across more than 45 families hosted on Hugging Face. It uses metadata analysis, tokenizer similarity, and weight-level signals to build fingerprints and operates in two modes: Compare and Scan. Cisco reports 100% recall on standard derivatives and cross-organizational fine-tunes. The company made the tool freely available on GitHub under the cisco-ai-defense repository to encourage industry-wide adoption of provenance standards.
Source: Read the original article
