人工智能安全测验

当您参加测验时，就会出现答案和解释。

1. 目前正在评估一个公共聊天室和一个具有写权限的内部智能体。为什么需要单独的威胁模型？
   • 所有AI应用程序都具有相同的信任边界
   • 它们的资产、权限、对手和可能带来的后果各不相同
   • 写权限会自动使内部智能体变得安全
   • 公共系统无法暴露私有数据
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