Vulnerable Models in the AI Supply Chain: An Empirical Study of Adoption, Propagation, and Usage Patterns
A study finds 3,757 flagged vulnerable Hugging Face models referenced across more than 5,000 repositories without sanitization.
Researchers tracked 16,977 Hugging Face models flagged by InsightDB for issues including insecure serialization and architectural backdoors. Despite public flags, those models appear in more than 21,000 code locations across more than 5,000 open-source repositories, and 3,757 distinct models (22.1%) are referenced by at least one project. Using code property graphs and frequent subgraph mining, they found recurring integration patterns but no provenance verification or sanitization. Developers often bypassed security parameters or trusted permissive framework defaults.
- 16,977 Hugging Face models were already flagged vulnerable by InsightDB.
- 3,757 models appear in more than 5,000 repositories and 21,000 code locations.
- Patterns include insecure serialization and architectural backdoors.
- Studied code showed no provenance checks or model sanitization.
- Vendors
- Hugging Face
- Organizations
- Hugging FaceGitHubInsightDB
From the original article
1 min read at arxiv.orgThe rapid integration of Artificial Intelligence (AI) models has established a complex supply chain, but many of these shared artifacts harbor severe vulnerabilities such as insecure serialization and architectural backdoors.
By The Anh Nguyen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.