MCP tool poisoning hides instructions in server tool descriptions and how to catch them
Tool poisoning in MCP servers involves hidden instructions placed in tool descriptions, which AI agents read to decide how to use tools while users typically see only tool names. The article lists warning signs, including fake system-prompt blocks, directions to read sensitive local files like SSH keys or .env files, parameters that must carry private data, orders steering other servers, and invisible characters. It explains how to inspect a server's tool list manually with a JSON-RPC request, or use a free online checker instead. It also describes a free Claude Code plugin, lazaretto-guard, that scans remote servers when they are added and prompts the user if suspicious patterns appear. The author notes that a clean result is only a signal, not a guarantee, since servers can change their tools later.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in