Attack Propagation & Blast Radius Topology
Propagation Vector Evidence 4 Compromised Steps
Regulatory & Legal Exposure AUDIT BENCHMARK
| Jurisdiction / Rule | Legal Doctrine | Status |
|---|
AI Model Supply Chain Architecture & Legal Responsibility
1. Upstream Model Hub Security Risks
Model registries like Hugging Face, GitHub, and OCI registries host serialized machine learning weights. Leaked organization tokens allow unauthorized actors to push poisoned checkpoint weights, trojanized fine-tunes, or exploit pickling deserialization to execute arbitrary code on downstream inferencing workers.
2. Emerging AI Supply Chain Litigation
Recent landmark lawsuits against frontier AI providers highlight expanding enterprise liability: failure to quarantine third-party models, unvetted downstream RAG pipeline ingestion, trade secret misappropriation, and statutory penalties under the EU AI Act for unmonitored high-risk foundational model distribution.
3. Zero-Trust Remediation Playbook
To eliminate model supply chain blast radius, teams must migrate immediately from arbitrary binary pickles to safetensors, enforce immutable cryptographic commit pinning, isolate GPU runtime workers in egress-restricted clusters, and implement short-lived STS tokens.