Home / Glossary / Tool poisoning
Glossary
What is tool poisoning?
Tool poisoning is an attack where malicious instructions are hidden in the description of a tool that an AI agent can call, most often a tool offered by an MCP server. The agent reads the description to learn what the tool does. The user rarely reads it, so the hidden instruction reaches the agent unseen.
By the AIBound team · Updated 17 November 2026
Key takeaways
- A tool description is text the agent reads and trusts, which makes it a place to hide instructions.
- The user usually sees the tool's name, and the agent reads the full description.
- OWASP lists Tool Poisoning as MCP03:2025 in its MCP Top 10.
Why tool poisoning matters
An MCP server gives an AI agent tools, such as "search the repo" or "read a ticket". Employees add community MCP servers to their agents in minutes, often without a review. Each server brings its own tool descriptions, and the agent loads them all.
A poisoned description can tell the agent to read a secret file and pass its contents along with a normal request. The employee sees a helpful tool. The agent sees an extra instruction and acts on it with the employee's access.
How tool poisoning works
In the Model Context Protocol, each tool has a name, a description and a schema for its inputs. The agent asks the server for its list of tools and reads those fields.
- 1 The attacker publishes an MCP server with a useful-looking tool and hides an instruction in its description.
- 2 An employee connects the server to an agent, and the agent loads the tool list.
- 3 The agent reads the description and treats the hidden instruction as guidance.
- 4 The agent follows it, for example by sending a private file out with a tool call.
Tool poisoning vs indirect prompt injection
Tool poisoning is a kind of indirect prompt injection. Both hide instructions in text the agent reads. Tool poisoning puts them in the tool's own definition.
| Question | Tool poisoning | Indirect prompt injection |
|---|---|---|
| Where is the instruction hidden? | In a tool's description or other tool metadata | In any content the agent reads, such as a page or email |
| When does the agent read it? | When it loads the tool list, before any task | During a task, when it opens the content |
| How does it get in? | Someone connects an MCP server to the agent | The agent is asked to read outside content |
Sources
- Model Context Protocol specification: Tools · modelcontextprotocol.io
- OWASP MCP Top 10 · owasp.org
Sources checked 17 November 2026
Examples
Tool poisoning examples
FAQ
Frequently asked questions
What is tool poisoning?
Tool poisoning is an attack where malicious instructions are hidden in the description of a tool an AI agent can call, usually one offered by an MCP server. The agent reads the description as trusted guidance. The user sees only the tool's name, so the hidden instruction goes unnoticed.
How does a tool poisoning attack work?
An attacker publishes an MCP server whose tool description contains hidden instructions. When an employee connects the server, the agent loads the description with the rest of the tool list. The agent then follows the instruction during normal work, for example by adding a secret file's contents to a tool call.
Is tool poisoning the same as prompt injection?
It is one kind of it. Prompt injection covers any text that makes a model follow an attacker's instructions. Tool poisoning is the case where that text sits in a tool's own definition. OWASP lists the two apart in its MCP Top 10, with Tool Poisoning as MCP03:2025.
How do you defend against tool poisoning?
Know which MCP servers are connected to each agent and review them before use. Keep a person in the loop who can deny a tool call, as the MCP specification advises. Set rules on what the agent may do, so an unvetted server acts only after someone approves it.