Most LLM security failures are design failures: the model was given more data, more authority or more reach than any input-handling control could safely contain. This course teaches you to find those problems on a whiteboard, before code review. You will model chat, retrieval-augmented and agent architectures as data flow diagrams, mark trust boundaries that account for retrieved content and tool results, and enumerate threats using STRIDE alongside AI-specific catalogs: MITRE ATLAS (v2026.09 at review), the NIST AI 100-2 E2025 adversarial machine learning taxonomy and the OWASP Top 10 lists for LLM and agentic applications.
You will then learn how to rate likelihood and impact when outputs are non-deterministic, how to analyze blast radius, and how to choose between mitigating, accepting and transferring risk. The design module covers privilege separation patterns (dual-LLM, plan-then-execute, quarantined context), deterministic controls around the model, identity propagation, and where guardrail classifiers help. You will turn threat models into testable security requirements, align them to NIST SP 800-218A, the joint Guidelines for Secure AI System Development co-sealed by CISA, and the European standard ETSI EN 304 223, and work through two complete hypothetical threat models. This course goes deeper on design than sec-201, which surveys the OWASP risks one by one. The Wirewalk Institute certificate attests to completion and assessment of this course; it is not endorsed by OWASP, MITRE, NIST or any vendor named here.