quiz bu lëkkale ak tegtal · Dëgër Niveau

Prompt Injection Detection Classifiers Quiz

Tests classifier scope, threshold tradeoffs, Prompt Guard 2 version limits, Microsoft Prompt Shields configuration, and defense-in-depth.

Laaj 1 de 8

What does an injection detection classifier return?

Xoolaatal laaj bu nekk8 laaj

Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.

  1. What does an injection detection classifier return?
    • A label or score about whether text resembles an attack under its policy
    • A guarantee that the input is safe
    • A corrected system prompt
    • Proof that the main model would obey the text
  2. What kind of text does Meta’s Prompt Guard 2 model card target?
    • Only generated model outputs
    • Explicit attempts to override prior instructions, labeled benign or malicious
    • All false statements and factual errors
    • Every harmful image
  3. What context-window limit does the Prompt Guard 2 card list?
    • One million tokens for every detector
    • A limit only on image inputs
    • 512 tokens for the documented models
    • No input limit
  4. Which risk comes with setting an injection-classifier threshold too aggressively?
    • Novel attacks will always be blocked
    • The system stops needing evaluation
    • All benign content is classified correctly
    • Legitimate security discussions may be flagged
  5. What can happen with a threshold that is too permissive?
    • Novel or reworded attacks may be missed
    • Every benign input is blocked
    • The classifier turns into a language model
    • The tool permissions change automatically
  6. What does Microsoft document about Prompt Shields configuration?
    • It guarantees no false positives
    • Intervention points must match where the attacks appear
    • The feature scans every deployment automatically
    • It cannot flag document attacks
  7. Why should a team record classifier version and intervention point?
    • To avoid tracking false positives
    • To bypass all model limits
    • To interpret and reproduce why particular content was scored or filtered
    • To make the score a legal conclusion
  8. How should classifier results fit into a security design?
    • As the sole proof that all unflagged text is safe
    • As a replacement for testing
    • As an automatic authorization for tools
    • As one signal combined with provenance, permissions, validation, and monitoring