引導式測驗 · 硬 等級
MCP 安全與工具中毒測驗
測試 MCP 中的工具中毒、拉扯和混淆副攻擊,以及定義固定和最小特權等防禦措施。
問題 1 的 8
MCP 中什麼是工具中毒?
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。
- MCP 中什麼是工具中毒?
- Overloading a server with requests
- Hiding malicious instructions in a tool's description that the model follows
- Corrupting a model's training data
- Deleting a tool from a registry
- 為什麼工具中毒往往對使用者來說是看不見的?
- Interfaces often show only a short name, not the full description
- Descriptions are encrypted
- Models delete descriptions after reading them
- MCP forbids showing descriptions
- 什麼是 MCP 地毯拉力?
- A server crashing during a call
- A user revoking a token
- A server changing its tool definitions after it was approved
- A client downgrading the protocol version
- 哪一種情況是混淆副攻擊?
- A server returns malformed JSON
- A model picks the wrong calendar tool
- A user types a password into chat
- Injected text in a public issue gets the agent to use its private repo access to leak data
- 為什麼僅安裝信譽良好的 MCP 伺服器還不夠?
- Reputable servers are always slower
- Injection can arrive through ordinary data returned by trustworthy tools
- MCP blocks reputable servers
- Reputation scores are not public
- What does hashing tool definitions at approval time let a client do?
- Detect later changes and ask for re-approval
- Speed up model inference
- Compress tool output
- Encrypt user prompts
- Which three conditions together make data theft especially likely?
- Fast network, large model and many users
- Open source, free tier and public registry
- Private data access, exposure to untrusted content and a way to send data out
- Long context, streaming and caching
- How does least privilege reduce MCP risk?
- It stops prompt injection completely
- 它使描述更短
- It removes the need for logging
- It limits how much damage a compromised or tricked server can do