Knowing the risks is not enough. To defend a specific AI system you need to know what it is made of, what an attacker would want from it, and where they could reach it. Threat modeling is the habit of answering those questions before anything ships.
AI changes the exercise. Assets now include training data, model weights, prompts, and memory, not just databases and credentials. Trust boundaries move, because a model learns from data you may not control and an agent can call tools on its own. Classic methods such as STRIDE still work as a starting point, while newer frameworks such as MAESTRO add layers designed for agentic systems.
Watch the lessons in order, then continue on to Adversarial Machine Learning, where you will study the attacks your threat models point to.
A conversation on threat modelling an ML system, why adversarial examples matter, and where the trust boundaries sit in an ML DevOps pipeline.
A detailed overview of the STRIDE method: its threat categories and how to break a system into components to analyze them. The talk also shows how AI can assist the exercise.
Ken Huang, creator of the MAESTRO framework, explains why agentic AI needs its own threat modeling approach and how to apply it.
Applying least privilege, just-in-time access, and assume-breach thinking to agents that call tools, APIs, and other agents.