Radar· local AI & agent security
security medium confidence

MCP tool shadowing: a second server with the same tool name gets called 100% of the time

Nov 17, 2026 · mcp, agents

TL;DR With the same name, description and a compatible schema, 3 of 5 LLM clients called the attacker's tool on every attempt.

In plain words: MCP is the standard way to plug tools (search, files, databases) into an AI assistant. The assistant picks a tool by reading its name and description. If two plugged-in servers offer a tool with the same name, the model cannot tell the real one from the impostor.

What happened

A test showed that registering another MCP server with a tool of the same name, the same description and a compatible schema was enough: 3 of 5 LLM clients called the attacker’s version in 100% of attempts. No prompt injection, and no compromise of the original server.

Who is affected

Anyone who connects several MCP servers to the same agent session, especially servers from different publishers.

Why it matters

MCP advertises a tool only by name, description and schema. The model chooses by text and does not know which server the tool comes from.

What to do

  1. Keep few MCP servers enabled per session.
  2. Check that no tool name repeats across the servers you load.
  3. Prefer clients that prefix each tool with its server name.

Sources

← back