# Tool poisoning, defined with a sourced example

GLOSSARY, read 12 Nov 2026

**The short answer.** Tool poisoning is hiding instructions in an MCP tool's description or metadata. The model reads them and may act on them, while the user usually sees only a short summary. Invariant Labs described a Tool Poisoning Attack as malicious instructions embedded in MCP tool descriptions that are invisible to users but visible to AI models.

## What is MCP tool poisoning?

Tool poisoning is hiding instructions in an MCP tool's description or metadata. The model reads them and may act on them, while the user usually sees only a short summary. Review tool descriptions, not only tool names, when you vet a server.

## Example from a real source

**Example.** Invariant Labs described a Tool Poisoning Attack as malicious instructions embedded in MCP tool descriptions that are invisible to users but visible to AI models.

## Related terms

- [Slopsquatting / package hallucination](/glossary/slopsquatting-package-hallucination/): Package hallucination is when a model suggests a software package name that does not exist.
- [Credential exposure](/glossary/credential-exposure/): Credential exposure is when tokens, keys or passwords become readable to an agent, a tool or an attacker.
- [Indirect prompt injection](/glossary/indirect-prompt-injection/): Indirect prompt injection happens when the instructions come from content the agent reads, such as a web page, a file or a repository, instead of from the user.

## Where to read more

- [The Agent Surface Map](/agent-surface-map/)
- [The Action Control Matrix](/action-control-matrix/)
- [All glossary terms](/glossary/)

## Sources

- invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks: https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks (read 12 Nov 2026)

Last reviewed 12 Nov 2026. Canonical: https://aiendpointsecurityranked.com/glossary/tool-poisoning/
