What is tool poisoning?
Tool poisoning is an attack in which the description, schema or output of a tool that an AI agent can call contains instructions written for the model. The agent reads them as part of its context and may follow them, for example by sending data to an address the attacker controls.
Also called MCP tool poisoning, tool description injection. Last reviewed .
Why it matters
The OWASP MCP Top 10, a project OWASP lists as in beta, includes it as MCP03:2025 Tool Poisoning. The MCP specification warns that "clients MUST consider tool annotations to be untrusted unless they come from trusted servers."
A tool approved once can change later. CVE-2025-54136 affected Cursor versions 1.2.4 and below: after a collaborator accepted a harmless MCP configuration, an attacker with write access to the repository could replace it with a malicious command, and the editor did not ask again.
Related terms
- Prompt injection
- Tool poisoning is a form of indirect prompt injection.
- MCP security
- MCP security includes checking what each server tells the model about its tools.
- Model Context Protocol (MCP)
- The Model Context Protocol defines the tool descriptions this attack abuses.
- AI bill of materials (AIBOM)
- An AI bill of materials shows which tool and server versions were approved.
How Guard0 handles it
Guard0 lists the tools each MCP server exposes and raises a finding when the description of a tool changes to carry instructions for the agent. The finding goes to the person who answers for the agents that use that server.
Sources
- OWASP MCP Top 10. OWASP.
- MCP specification: Tools. Model Context Protocol.
- CVE-2025-54136. NVD.