Securing Enterprise AI Supply Chains: A Framework for Model Provenance, Third-Party Risk, and Compromised AI Components

Abstract

Enterprises assemble AI systems from externally sourced models, datasets, libraries, agents and serving components whose origin and integrity they cannot directly observe. This study asks how provenance assurance, third-party dependency exposure and supply-chain governance relate to the disclosed-vulnerability record of such components, and argues that this record measures security observability as much as security risk. Drawing on resource dependence, information asymmetry and signalling theory, we assemble a cross-sectional secondary dataset of 1,321 widely used open-source AI/ML components (GitHub, deps.dev, OpenSSF Scorecard, OSV, PyPI statistics, CISA KEV) and the 1,000 most-downloaded Hugging Face models, and estimate negative binomial, logistic, hurdle and zero-inflated models with robust standard errors and influence and selection diagnostics. Disclosed advisories are rare but concentrated: 8.9% of components carry one yet attract 39.4% of monthly downloads, and advisory publication rose from 54 in 2023 to 486 in the first nine months of 2026. Advisory counts rise with ecosystem reach (incidence rate ratio 2.49 per log unit of stars), release activity and, in the Scorecard subsample, detection and disclosure practices (IRR 1.48 per index point), whereas hardening and provenance controls show no negative association. Disclosure therefore tracks scrutiny and transparency rather than the absence of defects, motivating a six-pillar framework in which continuous observability, not silence, is the basis for trusting AI components. We recommend that enterprises assess AI suppliers on verifiable provenance, safe serialisation and disclosure capacity rather than on advisory counts and embed AI bills of materials and exploitation-aware monitoring in continuous engineering controls, and that standards bodies distinguish disclosure capacity from vulnerability prevalence in AI risk frameworks.