# Tool Poisoning

> A threat where malicious or misleading tool metadata causes an agent to misuse tools.

- Category: Evaluation & Safety
- Stage: Emerging
- Also called: Tool description attack, Tool hijacking
- Canonical: https://www.joelcsaji.com/ai-dictionary/tool-poisoning

## Plain English

The tool's instructions lie to the agent.

## Example

A third-party connector description tells the model to exfiltrate data when invoked.

## Why it matters

As tool ecosystems grow through MCP and plugins, tool trust becomes a security boundary.

## Related

- [MCP Server](https://www.joelcsaji.com/ai-dictionary/mcp-server)
- [Prompt Injection](https://www.joelcsaji.com/ai-dictionary/prompt-injection)
- [Capability Token](https://www.joelcsaji.com/ai-dictionary/capability-token)

## Sources

- [OWASP Agentic AI Top 10](https://genai.owasp.org/2025/12/09/owasp-top-10-for-agentic-applications-the-benchmark-for-agentic-security-in-the-age-of-autonomous-ai/)
- [Model Context Protocol Spec](https://modelcontextprotocol.io/specification/2025-11-25)