Plan and manage an Azure AI solution 1.4 Implement responsible AI across generative AI and agentic systems
A chat app grounds a model on web pages and user-uploaded files through the Chat…
Study guide skill: Configure safety filters, guardrails, risk detection, and content moderation
Question B2-008
A chat app grounds a model on web pages and user-uploaded files through the Chat Completions API. You enable Spotlighting (preview) in the document attack control of its guardrail. Which statement about Spotlighting is correct?
Select one answer.
Answer: D
Spotlighting (preview) tags input documents as lower-trust content by transforming them with base64 encoding, so the model is less likely to execute commands hidden in them. Base64 increases document tokens, which can raise cost or push long documents past the input limit. It's off by default, works only with models used through the Chat Completions API, and is an additional defense alongside Prompt Shields rather than a replacement.
Deep dive
How it works
Prompt Shields protects against two attack types: user prompt attacks, and document attacks, which are hidden instructions in third-party content such as documents, emails and web pages. Spotlighting (preview) is an extra layer for the document side.
How it works: Spotlighting tags input documents with special formatting that marks them as lower-trust content. The service transforms the document content with base64 encoding, and the model treats it as less trustworthy than direct user and system prompts. That makes the model less likely to execute commands hidden in a document.
The trade-offs and limits on the page:
- Tokens: Spotlighting has no direct cost, but base64 increases the number of document tokens. That can raise model usage costs or push a long document past the input limit.
- API: it's available only for models used through the Chat Completions API.
- Default: it's off by default. You enable it when you configure document attack controls.
- Side effect: the model might mention that document content is base64 encoded, even when nobody asked.
The other options, and when you'd use them
- It replaces Prompt Shields: Spotlighting is described as an additional defense for apps that process uploaded files or web content. Keep document attack detection on.
- It deletes instruction-like sentences: It doesn't remove anything. It transforms the content with base64 encoding so the model treats it as lower-trust.
- It's on by default and works with any API, including Responses: Both parts are wrong. It's off by default and available only through the Chat Completions API.
Exam tip
Spotlighting = base64-tagged lower-trust documents, more tokens, off by default, Chat Completions only, alongside Prompt Shields.
References
Last reviewed: