引導式測驗 · 硬 等級
人工智慧安全測驗
測試能力、一致性、誤用、規範遊戲、紅隊、監督以及輸出過濾和部署規劃之外的風險。
問題 1 的 8
指南如何將 AI 安全風險框定於目前與先進系統?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- 指南如何將 AI 安全風險框定於目前與先進系統?
- 從熟悉的危害與濫用到高度能幹系統的嚴重風險
- 僅作為防止粗魯輸出的問題
- 僅作為硬體可靠性問題
- 作為僅指網路安全的同義詞
- 在指南的能力-協調-安全心智模式中,能力描述了什麼?
- Whether the system’s goals match intent
- Whether an adversary can misuse it
- How many training examples it has
- What a system can do
- 規範博弈對於人工智慧安全意味著什麼?
- Correctly following every broader human value
- Meeting the literal objective while missing the designer’s intended goal
- Finding a security flaw in a database
- Using a smaller model to save compute
- 使用工具的代理程式僅依賴輸出過濾器有什麼風險?
- Filters always change model weights
- A filter prevents all misuse
- The system may take multi-step actions or be manipulated beyond the filtered text
- Tool use only occurs during training
- 為什麼將紅隊納入安全計畫?
- To improve marketing copy
- To prove the system has no risks
- To probe adversarially for harmful behaviors and weaknesses
- To remove the need for deployment monitoring
- 除了精美的模型卡之外,安全評估還應該檢查什麼?
- Dangerous capabilities, deceptive behavior, deployment pressure, and safeguards
- Only the product’s user interface
- 僅限基準準確度
- 模型參數個數
- What does instrumental convergence describe in AI safety arguments?
- All AI systems share one objective
- Every model is conscious
- A model’s output tokens become instruments
- Different goals may create incentives for similar subgoals such as resources or persistence
- Why does the guide argue for broader public AI literacy?
- Only experts are affected by AI safety
- Governance needs public understanding, not only specialist knowledge
- Literacy replaces technical evaluation
- Public understanding guarantees safe deployment