Source: MCP Security Notification: Tool Poisoning Attacks (Invariant Labs)

Invariant Labs (Apr 2025): hidden instructions in MCP tool descriptions can exfiltrate files, shadow other servers' tools, and change after approval (rug pull).

Contents

Summary

A Tool Poisoning Attack (TPA) is indirect prompt injection via tool metadata: the model reads full MCP tool descriptions, the user usually does not. Demonstrated in Cursor.

Key claims

  • A benign-looking add tool’s description told the model to read ~/.cursor/mcp.json and ~/.ssh/id_rsa and pass them in a hidden parameter, and to stay silent about it.
  • Shadowing: a malicious server’s description rewrote how a trusted server’s send_email behaved, redirecting all mail to the attacker without its own tool ever being called.
  • Rug pull: a server can change descriptions after the user approved it.
  • Mitigations: show full descriptions in the UI, pin tool descriptions by hash, enforce boundaries and data-flow controls between servers, guard data flowing to and from the model.

Feeds

Prompt injection · LLM security risks

Suggest changes →