Question

A Business Solutions Architect is hardening a Copilot Studio customer-service agent that grounds its responses on documents uploaded by external partners to a SharePoint library. During a security review, the team discovers that a malicious partner embedded hidden instructions inside an uploaded PDF that could cause the agent to reveal internal pricing rules when a customer asks an unrelated question. Which mitigation most directly addresses this specific indirect prompt injection vector?

Accepted Answer

Establish a content validation and sanitization pipeline that screens uploaded grounding documents before they are indexed, and constrain the agent with system instructions that treat grounding content as untrusted data rather than instructions

Answer

Enable authenticated sign-in so only registered customers can start conversations with the agent

Answer

Apply a Data Loss Prevention (DLP) policy that blocks the SharePoint connector from being combined with premium connectors

Answer

Increase the model temperature so responses are more varied and harder to manipulate deterministically