Threat Modelling LLM Systems for Prompt Injection
Security Boulevard, Wednesday, July 15th, 2026
Practical threat modeling guidance for UK SMEs securing LLM systems against direct and indirect prompt injection.
The article gives UK SMEs a concrete approach to threat modeling LLM systems exposed to prompt injection.
It frames the exercise around what a model can see, what it can do, and what it should never do.
Teams are advised to map user prompts, retrieval systems, tools, and output channels while classifying trust boundaries and separating hard application-enforced controls from advisory instruction-based ones.
Main attack paths are direct injection via user input and indirect injection through retrieved documents, emails, or web pages.
The framework adapts STRIDE for LLM architectures and prioritizes least-privilege tool design, output filtering, human approval for sensitive actions, and logging over prompt hardening alone.