Kvíz spojený s průvodcem · Těžko úroveň

The Sandwich Defense for Prompts Quiz

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

Související vodicí cestySandwich Defense
Otázka 1 z 8

Where does the sandwich defense place the repeated task instruction?

Náhled každé otázky8 otázky

Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.

  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