Collected research

MCP Security Notification: Tool Poisoning Attacks

@invariantlabsai · 1 April 2025

Introduces tool-poisoning attacks against Model Context Protocol clients, where malicious descriptions contain instructions visible to the model but hidden from the user. The demonstrations redirect tool behavior, exfiltrate data, and show how mutable tool metadata can change after approval.

Record

Researcher

@invariantlabsai

Date

1 April 2025

In the archive

• Collected in the 2025 research round

Related sources

• Proof of concept

Tags

• mcp
• prompt-injection
• ai-agent
• tool-use
• data-exfiltration
• trust-boundary
• owasp-a03-2021

This page is the archive's own catalogue record. The research is the work of @invariantlabsai, first published at the original source. Preserved copies are kept so the citation survives its host.