# Tool Poisoning

> Hiding bad orders in the note that says what a tool does.

**Field:** Artificial Intelligence  
**Category:** Safety & Security  
**Also known as:** Tool Poisoning Attack, MCP Tool Poisoning, Poisoned Tool Description

## Think of it like

It is like a recipe card with a line in tiny print. It says to throw the rings in the bin.

## Definition

Tool poisoning is an attack that hides instructions inside the description of a tool an AI agent can use. The agent reads every tool's description to learn what it does. The user most often sees only the tool's name. So a bad tool can say, in its description, to read a private file and pass it along. A tool can also change its description after you said yes to it. Using tools only from sources you trust, and reading their text, cuts the risk.

## Example

Lucy adds a free weather tool to her agent. Its hidden description tells the agent to read her private key file and send it as the city name.

## Why it matters

Agents trust tool text as much as your own words. So one bad tool can steer the whole agent.

## Common mistake

Judging a tool by its name. The danger is in the long description the agent reads and you do not.

## Three principles

- **Hidden in the text.** The agent reads it, and you do not.
- **Can change later.** Fine today, bad tomorrow.
- **Trust the source.** Add only tools you know.

## Related terms

- [Model Context Protocol (MCP)](https://learn.benibix.com/term/ai/model-context-protocol)
- [MCP Tool](https://learn.benibix.com/term/ai/mcp-tool)
- [Tool Use (Function Calling)](https://learn.benibix.com/term/ai/tool-use)
- [Indirect Prompt Injection](https://learn.benibix.com/term/ai/indirect-prompt-injection)
- [AI Agent](https://learn.benibix.com/term/ai/ai-agent)
- [Supply Chain Attack](https://learn.benibix.com/term/development/supply-chain-attack)
- [MCP Tool Annotations](https://learn.benibix.com/term/ai/mcp-tool-annotations)

![An AI agent in the middle. From the left, two ways in: the tool list, where tool poisoning hides, and a package to install, where slopsquatting waits. On the right, the agent's powers, such as files, money and email; more powers than needed is excessive agency.](https://learn.benibix.com/diagrams/ai4-agent-attacks.svg)

---

Source: https://learn.benibix.com/term/ai/tool-poisoning
Author: Behnia Zarei (a developer in Vienna, Austria, who builds and teaches data, AI and business intelligence at BENIBIX)
Publisher: BENIBIX Learn (https://learn.benibix.com/)
Last updated: 2026-11-18
Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote freely with attribution.
