Kvíz spojený s průvodcem · Těžko úroveň

Kvíz o bezpečnosti MCP a otravě nářadím

Testuje otravu nástrojem, přetahování koberců a útoky zmatených zástupců v MCP, plus obranu, jako je připnutí definice a nejmenší privilegia.

Související vodicí cestyMcp zabezpečení a otrava nářadím
Otázka 1 z 8

What is tool poisoning in MCP?

Náhled každé otázky8 otázky

Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.

  1. What is tool poisoning in MCP?
    • 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
  2. Why is tool poisoning often invisible to users?
    • Interfaces often show only a short name, not the full description
    • Descriptions are encrypted
    • Models delete descriptions after reading them
    • MCP forbids showing descriptions
  3. What is an MCP rug pull?
    • 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
  4. Which scenario is a confused-deputy attack?
    • 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
  5. Why is installing only reputable MCP servers not enough?
    • Reputable servers are always slower
    • Injection can arrive through ordinary data returned by trustworthy tools
    • MCP blocks reputable servers
    • Reputation scores are not public
  6. 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
  7. 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
  8. How does least privilege reduce MCP risk?
    • It stops prompt injection completely
    • It makes descriptions shorter
    • It removes the need for logging
    • It limits how much damage a compromised or tricked server can do