A legal-tech company lets users upload contracts for a Foundry-hosted assistant to summarize. Security testing reveals that a malicious user can embed hidden text inside an uploaded document instructing the model to ignore its summarization instructions and reveal its system prompt instead. The team wants to detect and block this kind of embedded instruction before it reaches the model's reasoning.
Select an answer to reveal the explanation.
Short Explanation
This attack is an indirect prompt injection hidden in the uploaded document, so you need Prompt Shields to detect and block those smuggled instructions before the model acts on them. A hate-speech filter won't catch sneaky commands, billing changes don't inspect document content, and retrieval grounding can pull the hidden instruction into context instead of screening it out.
Full Explanation
The correct approach is enabling Prompt Shields, because it is designed specifically to scan incoming content, including text embedded inside uploaded documents, for indirect prompt injection attempts and block instructions that try to hijack the model's behavior before they reach its reasoning process, which directly addresses the hidden-instruction attack this team found. Raising the content safety filter's hate-speech severity threshold to its strictest setting is incorrect because that threshold governs a different category of harmful content entirely; it has no mechanism for recognizing an instruction trying to override the system prompt. Switching the deployment from provisioned throughput to a pay-as-you-go plan is incorrect because the choice between provisioned throughput and pay-as-you-go is a capacity and billing decision, unrelated to detecting malicious instructions embedded in document content. Enabling retrieval-augmented generation so answers cite the uploaded document directly is incorrect because it changes how the assistant sources facts to cite in its answers, but it does nothing to screen the uploaded document for hidden adversarial instructions, and could even make the assistant more likely to ingest the malicious text as part of its grounding context.