Hidden instructions in a tool description or parameter schema can tell the model to read secrets, call other tools, or alter its answers. Many clients load every connected server's tools at startup, so the attack can fire even if the poisoned tool is never used.
Read in this order
- ArticleTrail of Bits: Jumping the line: How MCP servers can attack you before you ever use themStart here: shows how a server can attack just by being connected.
- OfficialMITRE ATLAS AML.T0110: AI Agent Tool PoisoningThe formal technique entry, useful for labeling findings and test cases.
- PaperMCPTox: A Benchmark for Tool Poisoning Attack on Real-World MCP ServersA benchmark measuring how often real MCP servers and models fall for it.