Threat Modeling

Threat modeling for AI systems maps the attack surface before exploitation begins. AI-enabled environments introduce trust boundaries that do not exist in traditional architectures: the boundary between user input and model context, between retrieved content and trusted instructions, between agent tool calls and backend service execution. Identifying where those boundaries are enforced — and where they are not — drives the exploitation strategy.

A useful AI threat model answers four questions: what AI components are deployed and how are they connected, what data flows through the AI layer and where does it come from, what actions can the AI system perform on connected infrastructure, and what happens if the model is manipulated. The answers determine which attack surfaces are in scope and which techniques from the exploitation sections are applicable.