Threat modeling workshop
Threat modeling workshop for agents, RAG and MCP
Together we model one concrete system, from the point where untrusted content comes in to the last tool call. This format fits while the system is still being built and we can define the boundaries before they get built.
- Scope
- One concrete system, usually an agent, a RAG application or an MCP server. We agree on the scope in the intro call.
- Effort
- 1 to 2 days, with your engineering team. You receive the proposal after the intro call.
- Outcome
- A threat model your team keeps maintaining and a list of the places that need a closer look.
Who the workshop is for
Your team builds an agent, a RAG application or an MCP server and is still at the design stage or in the middle of the build. Or an enterprise customer asks in a security questionnaire for a threat model you can show.
When another format fits better
- The system goes live soon. A workshop still helps, but it does not check whether the code that was built matches the threat model. The go-live review with targeted tests against staging fits better.
- Your application has no agent. Threat modeling for applications and APIs is part of application security. The process is similar, the questions are different.
- You need a seal of approval or a legal opinion. You can show a threat model to customers. It is not a seal of approval, and whether NIS2 or the AI Act applies to you is for your lawyers to decide.
Contents
What we cover in the workshop
-
Data flows and trust boundaries
We draw where untrusted content comes from, which data the agent sees and which data goes to the model provider.
-
Permissions per tool
For every tool we record which data it reaches under which identity.
-
Lethal trifecta
We look for the places where private data, untrusted content and a way to send data out meet in the same context.
-
Approval points
We decide where the agent has to wait for a person and what the approval should show them.
-
Mapping to OWASP Agentic
We map the threats to the OWASP Top 10 for Agentic Applications 2026 where an identifier fits.
-
Places for a closer look
At the end there is a list of the places that need a closer look, in a review or by your team.
Process
How the workshop works
-
Intro call
We choose the system and clarify how far along it is and who from your team takes part.
-
Preparation
I read what already exists beforehand, such as architecture sketches, tool definitions or early code in the repository.
-
Workshop
One to two days with your engineering team, following the path from untrusted content coming in to the last tool call.
-
Threat model and list
You get the threat model in a form your team keeps maintaining, and the list of places that need a closer look.
If you book a go-live review for the same system afterwards, I credit half of the workshop toward it.
Writeups
Further reading
In these articles I describe patterns that this page is about, each with code or configuration.
Contact
Request a workshop
In the intro call we clarify which system suits the workshop and who from your team should join. You then receive a proposal.