Cisco has introduced a new open source tool called the Model Provenance Kit, which aims to help organizations manage the complexities and risks associated with third-party AI models. This toolkit responds to growing supply chain risks, such as AI model poisoning, biases, and unverified claims from developers, which can lead to security and compliance issues. The Model Provenance Kit generates a ‘fingerprint’ for each model, helping users track lineage and ensure models are suitable for their use cases. By offering functionalities to compare and scan models, Cisco’s tool is positioned as a proactive measure to enhance the integrity of AI deployments, particularly in light of recent accusations around unauthorized distillation from proprietary models. The open source kit is now available on GitHub, with a dataset of model fingerprints provided on Hugging Face.

Cisco: Cisco Systems is a leading multinational technology company focused on networking, cybersecurity, and AI infrastructure solutions. It has recently advanced its AI security offerings, including updates to the AI Defense platform and hosting events like the AI Summit. In this news, Cisco released the Model Provenance Kit, an open-source tool to help organizations verify the origins and lineage of third-party AI models, addressing security and compliance risks.
Hugging Face: Hugging Face operates as a central open-source platform for machine learning models, datasets, and tools, facilitating community collaboration on AI development. Recent reports highlight its role in hosting diverse models amid challenges like varying maintenance of model cards and metadata. The news references it as a primary repository for third-party models that may lack verified provenance, underscoring the need for tools like Cisco’s kit.
Model Provenance Kit: The Model Provenance Kit is a Python-based open-source toolkit and command-line interface developed by Cisco for AI model provenance verification. It generates fingerprints from metadata, tokenizer similarity, and weight-level signals to compare models or scan against a database of base model fingerprints hosted on Hugging Face. This tool directly tackles issues like obscured lineage from fine-tuning and repackaging, enabling better incident tracing and risk mitigation.

`json
{
“Supply Chain Risks”: “Concerns over AI model poisoning, biases, and unverified developer claims are prompting new tools for lineage tracking in enterprise deployments.”,
“Provenance Standards”: “Emerging practices like model fingerprinting are influencing approaches to model verification.”
}
`