A company builds a chatbot using Azure OpenAI with a system prompt containing proprietary business logic. The security team is concerned that users could extract this system prompt through carefully crafted questions. Beyond Prompt Shield, which Azure OpenAI feature can be configured to add a layer of protection against system prompt extraction?

The correct approach is the Assistants API with server-side system prompt storage that is never exposed to the client. The Azure OpenAI Assistants API manages the assistant's instructions, which are the system prompt equivalent, server-side within the Azure OpenAI service, not in client-visible API request bodies. Users and client applications never see the system instructions since they are managed through the Assistants API and referenced by Assistant ID, making extraction through prompt manipulation significantly more difficult. The content filtering approach with jailbreak detection at high severity to redact assistant instructions from model context is incorrect because content filtering with jailbreak detection helps block obvious extraction attempts, but sophisticated prompt manipulation may still succeed in extracting system prompt content. The lower context window approach is incorrect because limiting the context window does not prevent system prompt extraction; the system prompt is still accessible to the model and can be recalled. The fine-tuning approach to embed business logic into the model weights instead of a system prompt is incorrect because fine-tuning embeds behavioral patterns, but proprietary business logic cannot be reliably or completely transferred to model weights; some system prompt remains necessary.