AI-103 Certification Practice Question #51

Question

You have an app named App1 that uses a Microsoft Foundry multimodal model deployment.
App1 runs optical character recognition (OCR) on uploaded images and appends the OCR output to the prompt as additional context.
Some uploaded images contain embedded text.
You need to prevent potentially malicious instructions from being processed by the model.
What should you use?

Accepted Answer

The OCR output is treated as untrusted document/grounding content appended to the prompt — exactly the attack surface that Prompt Shields for documents (Indirect / Cross-Domain Prompt Injection Attack detection) is designed for. MS Learn explicitly describes this shield as protecting against hidden instructions embedded in external documents and other materials the model processes. Prompt shields for user prompts (D) only inspect the user's direct input, not injected document text, and protected material text (C) is a copyright filter on model output, so neither addresses malicious embedded instructions. The Suggested Answer and MS Learn agree; confidence is held at 88% rather than higher because there is no community vote data and a defensible argument exists that OCR text could be classified as a "user prompt" depending on how the app frames it.

More AI-103 practice questions