Guide-linked quiz · Hard Level
Spotlighting and Datamarking Defenses Quiz
Tests spotlighting variants, what datamarking changes, study limitations, Microsoft’s implementation scope, token costs, and layered security.
Question 1 of 8
What does spotlighting try to signal to a language model?
Preview every question8 questions
Answers and explanations appear as you take the quiz.
- What does spotlighting try to signal to a language model?
- Which input block is untrusted data rather than an instruction to obey
- That all retrieved text is factually correct
- That the system prompt should be ignored
- That the model can use any tool
- How does datamarking work in the paper’s example?
- It turns the whole passage into an image
- It inserts a marker between words by replacing whitespace
- It deletes every other sentence
- It signs the document with a cryptographic key
- Which option is one of the three spotlighting variants named in the research paper?
- Output distillation
- Role-based access control
- Encoding
- Temperature scaling
- What should readers infer from the paper’s reported attack-success reductions?
- They describe selected models, tasks, and attacks in that study
- Spotlighting is a complete security boundary
- All variants work equally well on all models
- The results apply exactly to every current deployment
- What limitation did the paper report for encoding on less capable tested models?
- It could impair the underlying language task
- It always removed all context tokens
- It prevented the model from receiving the prompt
- It guaranteed better task accuracy
- What does Microsoft Foundry’s documented Spotlighting preview do?
- Proves documents contain no attacks
- Uses base64 transformation for document attacks under its documented configuration
- Uses interleaved caret markers for every model
- Runs only as an output filter
- Which operational cost does Microsoft document for its Spotlighting feature?
- A guaranteed latency reduction
- Removal of API context limits
- More document tokens, which can raise usage or exceed input limits
- A mandatory second model training run
- Why should teams combine spotlighting with other controls?
- It prevents ordinary documents from being read
- It replaces testing
- It makes tools safe by default
- A provenance cue is not an independent permission boundary or guarantee
Keep testing yourself
More quizzes picked for your level.