Red teaming tests the system you built. But most AI systems are not built from scratch: they rely on pre-trained models, public datasets, open-source libraries, and third-party services. Every one of those is a link in the AI supply chain, and a weakness in any link becomes a weakness in your product.
This module covers two connected problems. The first is supply chain risk: malicious or tampered models downloaded from public hubs, poisoned dependencies, and compromised build pipelines. The second is data security: training and fine-tuning data that can be poisoned, and sensitive information that can leak through prompts, retrieval pipelines, and model outputs.
Watch the lessons in order, then continue on to Guardrails & Defensive Controls, where you will look at the controls that reduce these risks.
Checkmarx researchers show how malicious models can be built and published to public model hubs, how harmful code can be hidden inside model files, and how poisoning can be aimed at specific victims.
A SANS session that maps what a GenAI application depends on, from models and data to libraries and pipelines, and how to monitor and defend each part.
A short introduction to how attackers corrupt training or fine-tuning data, what that does to a model, and why it is hard to spot.
IBM Technology explains how sensitive data leaks through training data, prompts, RAG pipelines, and agents, and why traditional data protection tools miss it.