Guide-linked quiz · Hard Level

The Sandwich Defense for Prompts Quiz

Tests how prompt sandwiching is structured, its hypothesis, model-dependent limits, benchmark scope, and need for independent controls.

Related guide pathsSandwich Defense
Question 1 of 8

Where does the sandwich defense place the repeated task instruction?

Preview every question8 questions

Answers and explanations appear as you take the quiz.

  1. Where does the sandwich defense place the repeated task instruction?
    • Before and after untrusted content
    • Only inside the retrieved document
    • After the model’s answer
    • Only in an evaluation report
  2. How could a closing reminder help in this pattern?
    • It forces all models to obey the final sentence
    • It may refocus the model on the task after the untrusted text
    • It creates cryptographic authorization
    • It proves the page is safe
  3. What does sandwiching change in the application?
    • The user’s authorization level
    • The retrieved document’s source
    • The prompt template’s text structure
    • The model’s trained weights
  4. Why is the sandwich defense not a security boundary?
    • It does not use punctuation
    • It always hides the external document
    • It automatically blocks tool calls
    • The reminder and the hostile passage remain text in the same model context
  5. What did the cited domain-camouflage preprint evaluate?
    • Several prompt-level defenses across three model families and synthetic professional-document domains
    • Every commercial model and real enterprise document
    • Only a product’s password reset flow
    • Every possible prompt injection attack
  6. What does the preprint imply about effectiveness across models?
    • Model choice has no effect
    • It can vary by model and no tested prompt defense eliminated every threat in weaker models
    • The defense works identically on all models
    • Sandwiching guarantees zero attacks
  7. Which control addresses whether an agent is allowed to perform an external action?
    • A longer source document
    • A higher confidence statement
    • Application-side tool permission checks
    • A repeated closing sentence alone
  8. How should a team decide whether sandwiching helps its application?
    • Assume it works because it is easy to add
    • Measure only answer length
    • Skip task-quality checks
    • Test it against a baseline on the same representative benign and malicious inputs

Keep testing yourself

More quizzes picked for your level.

View all quizzes