Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ
Bài kiểm tra tiêm nhắc gián tiếp
Nhận biết các cuộc tấn công do nội dung không đáng tin cậy thực hiện và áp dụng các biện pháp kiểm soát theo lớp xung quanh luồng dữ liệu, công cụ và các hành động tiếp theo.
Câu hỏi 1 của 8
What makes a prompt injection indirect?
Xem trước mọi câu hỏi8 câu hỏi
Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.
- What makes a prompt injection indirect?
- The user types the instruction directly into chat
- The model receives no text
- The adversarial instruction arrives inside content the assistant is asked to process
- The instruction is part of the system prompt
- Which content could carry an indirect prompt injection?
- A webpage, email, uploaded file, or retrieved document
- Only a developer instruction
- Only a model’s weights
- A hardware temperature reading
- What can an indirect injection try to cause an agent to do?
- Improve hardware throughput
- Change the user’s password without access
- Alter a file format automatically
- Change behavior or misuse a connected tool, including exposing data
- Why does tool access raise the risk of prompt injection?
- Tools make prompts deterministic
- A manipulated response may trigger actions or data sharing beyond the text answer
- Tools remove untrusted content
- Tool calls cannot have side effects
- Which is a defense-in-depth measure?
- Rely only on a prompt saying “ignore attacks”
- Give every agent access to all data
- Restrict tool permissions and validate arguments outside the model
- Remove action confirmations
- Why use structured outputs for data extracted from an untrusted document?
- They prove the document is trustworthy
- They can limit free-form data flow and make fields easier to validate
- They prevent all attacks automatically
- They grant the model more permissions
- Does a model’s refusal behavior alone make a connected workflow secure?
- No; authorization and tool controls should be enforced outside the model too
- Yes, for every model and attack
- Yes, if the prompt is long
- No, because models cannot refuse
- What should a user review before confirming a consequential agent action?
- Only the agent’s tone
- Only the number of tools called
- Nothing if the content came from a webpage
- The exact action, destination, and information being shared