Securing the AI Supply Chain

This is your next chapter at TryTrainMe. In the Supply Chain Attack Vectors room, your investigation confirmed the breach: a malicious pickle, a fake repository, and dependency confusion. Your work impressed the board. You have been promoted to Security Engineer and given the mandate to build an internal supply chain security testing lab (SupplySecLab) so nothing like this gets through again.

SupplySecLab closes each gap from that incident:

• No policy governed which model formats were acceptable; in Task 2 you will see how to address this
• The model's integrity was never verified before deployment; Task 3 shows you how to change that
• The model was never scanned before it entered the pipeline; Tasks 4 and 5 walk you through the tools that catch this
• Hidden logic inside the model's architecture went undetected; in Tasks 6 and 7 you will learn how to find it
• A malicious package slipped through because dependencies were never audited; Task 8 covers how to prevent this
• The production system ran on an external prompt that was never reviewed; Task 9 shows you how to assess and govern this

Learning Objectives

• Use SafeTensors and weights_only=True to eliminate the pickle-based code execution risks introduced in Supply Chain Attack Vectors
• Verify model integrity using SHA-256 checksums and model card review
• Scan models with Fickling and ModelScan to detect malicious content before deployment
• Audit dependencies with pip-audit and generate Software Bills of Materials (SBOMs) with Syft
• Assess API providers against a supply chain security checklist and establish behaviour monitoring controls

Tasks in this room use a Linux VM for static analysis and AI Agents for live analysis and API provider assessment. 

Prerequisites

• Completed Understanding AI Supply Chains (supply chain concepts)
• Completed Supply Chain Attack Vectors (malicious models, dependency confusion, repository attacks)
• Recommended: AI/ML Security Threats (foundational AI security concepts)

Framework Alignment

This room maps to OWASP LLM03: Supply Chain Vulnerabilities (opens in new tab), MITRE ATLAS AML.T0010 (ML Supply Chain Compromise), and NIST AI RMF Govern 1.2, Measure 2.2, and Manage 2.1.