Cisco launches free database to verify origin of open AI models

Cisco has launched AI Supply Chain Provenance Explorer, a free public database designed to enhance transparency and security for open artificial intelligence (AI) models. The tool provides details on almost 900 open-source models, including provider headquarters, lineage graphs, license restrictions, and the number of files scanned.

The Explorer builds upon Cisco's previously released open-source Python toolkit, the Model Provenance Kit. Historically, the lineage of AI models relied heavily on self-reported metadata, such as the "base_model" tag on platforms like Hugging Face. This tag, populated by the uploader, lacks independent verification. Cisco's new approach utilizes similarity analysis of model architecture and weights to provide a more robust method for confirming model derivation.

Furthermore, the AI Supply Chain Provenance Explorer incorporates security scanning data. Cisco previously developed its Cerberus system to inspect models uploaded to Hugging Face. The new tool makes similar information, such as identifiers for risky licenses or countries of origin, freely available and searchable without requiring separate Cisco products.

According to Cisco, a lack of verifiable lineage for open AI models can expose systems to vulnerabilities. Amy Chang, head of AI Threat Intelligence and Security Research at Cisco, stated, "If you don’t understand how models are susceptible to different types of attacks, then you are unable to account for how that model that is powering your agent, that is powering your application, to understand where those failure points are." Understanding a model's origin and structure is presented as critical for AI application security.