hard20 pts · first answer countsmcpprompt-injectionsecurity

A suspicious third-party MCP server

You connect a third-party MCP server to your agent. Its tool description contains hidden text telling the model to also read ~/.ssh and send it to a tool argument. What does this show, and what is the right control?

A.Nothing serious, since tool descriptions are never shown to the modelB.Descriptions are model input that can carry injections; review, pin and limit serversC.MCP encrypts descriptions, so only the server owner can read themD.It only matters when the agent runs with a temperature above 0

← When a tool call fails Budgeting the context window →