Tool descriptions are prompts the model reads, so a server can change agent behavior without any change to code you reviewed. Scanners pull tool metadata from configured servers and flag poisoning, shadowing and changes between scans. Run them in a sandbox, since they may launch the very servers they inspect.
Read in this order
- ArticleMCP Security Notification: Tool Poisoning Attacks (Invariant Labs)Start here: the original tool poisoning demo and why descriptions are an attack surface.
- Open sourceSnyk Agent Scan (agents, MCP servers, and skills)A scanner you can run on your own agents, MCP servers and skills.
- ArticleJumping the line: How MCP servers can attack you before you ever use them (Trail of Bits)Shows how a server can attack during connection, before any tool call.