AI-103: Developing AI Apps and Agents on Azure

Azure AI Content Safety Prompt Shields analyzes two distinct attack surfaces: User Prompts, for direct jailbreak attempts typed by the user, and Documents, for indirect (cross-domain) prompt-injection attacks in which malicious instructions are hidden inside third-party content — such as documents, emails, or, as here, OCR text extracted from an image — that is appended to the prompt as additional context rather than authored by the user. Because the embedded text arrives as external content folded into the prompt, Prompt Shields for documents is the control designed to detect and block instructions hidden in it. Image moderation screens for visual harm categories, not instructions; protected material detection targets copyrighted content; and Prompt Shields for user prompts addresses direct user-typed jailbreak attempts, not third-party content.