Threat Modeling Setup for GenAI and Agentic Systems
Explore how to set up a comprehensive threat model for GenAI and agentic systems by identifying key assets, actors, trust boundaries, and data flows. Understand security objectives related to confidentiality, integrity, availability, and misuse resistance. Learn to map security enforcement points, evidence requirements, and residual risks to prepare effective AI security reviews.
The LLMOps evidence packet defined what signal exists so that an incident can be investigated after the fact. It never asked whether the system was attackable in the first place. That is a different discipline, with its own artifact. Security reviews for GenAI and agentic systems fail when the discussion starts as a list of prompt injection tricks. Start instead from a boundary and data flow model that states concretely what can cross into what.
Build a threat model architecture package carried from prior chapters: a scoped document that enumerates assets, actors, trust boundaries, and flows across model calls, the retrieval pipeline, MCP tools, the AI gateway enforcement layer, and delegated identity. Prior observability and evaluation artifacts show runtime behavior and output quality, but they cannot prove a boundary crossing is safe, authorized, or non-bypassable. That proof needs a clearly stated system context: where enforcement occurs, and what must be logged to investigate an incident. ...