AI-103: Developing AI Apps and Agents on Azure

Foundry Agent Service guardrails support four intervention points - user input, tool call, tool response, and output - and for agents (as opposed to models) the only available response action is Annotate and block; a plain Annotate-only action is not offered for agents, so any control meant to prevent harmful content from being returned must use Block. Because the screenshot's blob URL is produced by the ticketing tool and only becomes available to the agent as content returned by that tool, the guardrail must cover the tool response - along with user input, tool call, and output - so image content is screened wherever it can enter or leave the run, with Block enforcing that flagged harmful content is stopped rather than merely logged. For storage access, Azure AI Content Safety's documented requirement is that its resource identity hold the Storage Blob Data Contributor (or Owner) role to read a blob via a blobUrl; Storage Blob Data Reader is not a supported role for this scenario. The least-privileged working configuration is therefore a system-assigned managed identity granted Storage Blob Data Contributor, which avoids both storage account keys and an unrelated user-assigned queue role.