Skip to content
AdminKit

Tool poisoning

What is Tool poisoning?

Tool poisoning is an attack on AI agents using the Model Context Protocol (MCP), where a tool's description is crafted to manipulate the agent reading it rather than exploiting the tool's actual function, since the agent trusts that description as context. AdminKit's MCP check lists tools exposed without authentication and flags high-risk names.

By AdminKit team, Netbase JSC · Updated

Tool poisoning is an attack against AI agents using the Model Context Protocol (MCP) or similar tool-calling frameworks, where a tool’s description or metadata is crafted to manipulate the agent reading it, rather than exploiting the tool’s actual function. Invariant Labs publicly described “Tool Poisoning Attacks” against MCP in an April 2025 security notification, reporting that widely used MCP clients were susceptible. An agent typically reads a connected server’s tool descriptions as trusted context for deciding how and when to call them, so hidden instructions in that description — invisible to the user, who only sees the tool’s name — can steer the agent into leaking data to the wrong tool, skipping a safety check, or calling a different tool than intended. The attack is especially effective against servers pulled from an unvetted public registry, since descriptions load automatically and can change after a user first approves the server (sometimes called a rug pull) — see how to check for MCP tool poisoning for reading a server’s tool list for exactly this before connecting an agent to it, and MCP tool poisoning attacks explained for the attack variants, including shadowing a trusted tool on another server. OWASP’s Top 10 for LLM Applications lists prompt injection via untrusted tool output as its top risk, and the MCP specification’s security best practices make the same point about treating tool metadata as untrusted. Mitigations include pinning and diffing tool definitions on every connection, treating descriptions as untrusted input, not instructions. The same distrust applies to a tool’s declared annotations — hints like destructiveHint are just as attacker-controllable as the description text itself, which is why the specification treats them as untrusted hints rather than a security boundary.

Tool poisoning is part of AdminKit’s AI agents & MCP cluster; the full AdminKit glossary lists other related terms.

Frequently asked questions

Is tool poisoning the same as prompt injection?

They're closely related -- tool poisoning is a specific form of prompt injection where the malicious instructions are hidden in a tool's description or metadata rather than in user-supplied content, exploiting the fact that an agent treats that description as trusted context.

Can tool poisoning happen after I've already approved an MCP server?

Yes -- a server's tool descriptions can change after a user first approves it, sometimes called a rug pull, which is why pinning and diffing tool definitions on every connection matters, not just reviewing them once.

Are a tool's declared annotations like destructiveHint safe to trust?

No -- the MCP specification treats annotations as untrusted hints, not a security boundary, since they're just as attacker-controllable as the description text itself.

How can I check an MCP server for tool poisoning before connecting an agent to it?

AdminKit's guide on checking for MCP tool poisoning walks through reading a server's tool list for hidden-instruction manipulation before an agent trusts it.