The original title is "The AI Supply Chain: The Next Evolution of Third Party Risk"

Short summary

Traditional third-party risk management is insufficient for AI systems where trust extends to model weights, embeddings, and retrieval pipelines. Key threats include model provenance attacks via pickle serialization, RAG poisoning through indirect prompt injection, and compromised external inference APIs. The article recommends hardened infrastructure controls: cryptographic model signing, multi-stage RAG ingestion pipelines, and treating all external data sources as untrusted.

• •AI supply chain risk extends beyond vendor agreements to model weights, embeddings, and RAG pipelines
• •Four major attack vectors: model provenance, RAG poisoning, external APIs, and framework dependencies
• •Defense requires infrastructure hardening: model scanning, cryptographic signing, and untrusted-data pipeline controls

Generated with AI, which can make mistakes.