Demo: threat-modeling the Larkspur assistant
On This Lesson 3 topics
This lesson is planned. The outline below lists the topics it will cover. The full lesson has not been published yet.
In this lesson
threat-model the Larkspur assistant and turn findings into design controls. This prepares ARC-046-LAB1.
Draft outline details
Step by step
- Draw the data flow: customer or staff user, assistant service, retrieval over product, order, and policy content, model provider, tool gateway to order and refund services, logs.
- Mark trust boundaries: user to service, service to provider, tool gateway to internal systems.
- Planned mistake: rely on the system prompt ("never reveal card numbers"). A tester pastes a card number into chat and an injected instruction in a product review makes the assistant echo it. Fix: mask card patterns at the input filter and never log them, filter outputs, and keep card data out of retrieved content.
- List threats per boundary (injection, cross-customer leakage, over-powered refund tool, provider retention, log exposure) with a control and owner for each.
- Add residual risks and test cases to the assurance plan, and record the decisions.
- Review with the role-played security and privacy officer and confirm no payment card data can reach the model.
Check your understanding
What was the planned mistake?
Show answer
Relying on the prompt as the control instead of enforcing masking and filtering in the architecture.
Practice
Apply this lesson with the module's lab, exercises and quiz.
Marking a lesson complete is optional and you can undo it any time. Your progress stays in this browser.