AI 보안 퀴즈

퀴즈를 풀면 정답과 설명이 나타납니다.

1. A public chatbot and an internal agent with write access are being assessed. Why need separate threat models?
   • All AI applications have identical trust boundaries
   • Their assets, permissions, adversaries, and possible consequences differ
   • Write access makes the internal agent automatically safe
   • Public systems cannot expose private data
2. What does a prompt saying “follow security rules” replace?
   • Account isolation
   • Secret management
   • Nothing; authorization and permission limits still need enforcement in code
   • Output validation
3. Which example is a training-stage threat?
   • An attacker changes one inference-time image
   • A user guesses a password
   • A database rejects a valid query
   • Poisoned training material that changes learned behavior
4. A model writes a database query from a user request. What safeguard applies before execution?
   • Validate and safely handle the generated output at the software boundary
   • Assume the model understands database permissions
   • Use the output as a command without inspection
   • Rely only on a prompt asking for safe SQL
5. Why restrict a tool to only the permissions it needs?
   • It prevents every adversarial input
   • A model error then cannot silently expose broader capability
   • It eliminates security testing
   • It guarantees confidentiality of all logs
6. What must a successful refusal test not be confused with?
   • Evidence that one test case passed
   • One observed model behavior
   • Proof that the whole system is secure across different paths and modalities
   • A useful regression example
7. How should security testing be conducted?
   • By probing unrelated external systems
   • Only after a breach
   • Without documenting scope or ownership
   • On representative failure paths in an authorized environment
8. What should security telemetry balance?
   • Enough diagnostic value with minimized collection of sensitive content
   • Maximum prompt retention regardless of purpose
   • No event data at all
   • Full response logging for every user indefinitely