Exercise - Prompt shields

Why it matters

An agent that reads retrieved documents or email can be attacked by text no user typed, so injection detection has to cover both entry points.

What to practise

The exercise runs in the Foundry classic portal, in a hub-based project, and shows the earlier name Azure AI Foundry. On the Prompt Shields panel of Guardrails + controls › Try it out, run a jailbreak-style user prompt, then a product manual with an instruction hidden in it as a document, then both together. Watch that the result reports each attack against its own input, user prompt or document. The exercise uses fictional data and suggests testing with further synthetic prompts from your own threat model.

How it fits

Matching direct and indirect shields to where text enters is 4.2.2; the portal test is the interactive side of the Content Safety check that 4.2.2 also calls through the API. The document case is the untrusted-content screening in 4.2.4. Turning extra synthetic attacks into a repeatable suite is 4.2.8.