Architectural guardrails
On This Lesson 4 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
place guardrails in the architecture so they hold even when the prompt is bypassed.
Planned topics
- Prompt injection: untrusted text (a customer message, a web page, a document) tries to steer the model. Treat all model input as untrusted and output as untrusted.
- Data leakage: limit what the model can see by access-controlled retrieval using the end user's permissions, plus output filters.
- Excessive agency: give the assistant the minimum tools, read-only by default, with confirmation or human approval for actions like refunds.
- Put controls in code and infrastructure: input and output filtering, schema validation of tool calls, allowlists, rate limits, and scoped credentials. Prompts alone are not a control.
- Layer them: no single guardrail is relied on.
Draft outline details
Demonstration
a flow for the Larkspur assistant: user, input filter, retrieval with user's entitlements, model, output filter, tool gateway (allowlist, scopes, approval step for refunds), audit log.
Check your understanding
Why is a prompt instruction alone not a guardrail?
Show answer
The model can be manipulated or ignore it; real controls must be enforced outside the model.
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.