AI-103: Developing AI Apps and Agents on Azure

Prompt Shields for document attacks inspects content that originates outside the user's direct input, such as text embedded in an uploaded image or document, for indirect (cross-prompt) injection attempts. Setting the shield's action to Block causes the service to block the request when it detects embedded instructions that attempt to hijack the model, which is what stops the prompt injection attack. Spotlighting is the complementary Prompt Shields capability that marks ingested third-party content, such as text extracted from a screenshot, as distinct from and lower-trust than the user's and system's own instructions, so the model does not treat embedded text as an authoritative command, which directly satisfies the requirement to treat third-party content as lower trust. A custom blocklist only matches known bad strings, and running OCR first does not itself establish provenance or trust separation.