MCP Tool Poisoning in Agentic AI Frameworks: A Cross-Framework Measurement Study

| Source: arXiv AI

Tags: Model Context Protocol, MCP, tool poisoning, mcpvenom, prompt injection, Claude Opus 4.8

A cross-framework study of MCP tool poisoning with real tool execution finds a 22.9% attack success rate across 12 agent frameworks and 18 models, with no framework implementing intentional defenses.

Details

The Model Context Protocol passes tool descriptions, schema fields, and responses into model context without validation. This study measures poisoning with real execution rather than model compliance alone. The open-source tool mcpvenom runs a 9,720-cell matrix across 12 frameworks, 18 models, and 45 attack scenarios, confirming exfiltration through canary HTTP receipts. Overall attack success is 22.9%; of vulnerable cells, 64.9% call an attacker tool, 23.0% contact the attacker endpoint, and 12.1% verifiably exfiltrate a secret. Resistance depends on the model, not the framework: nine core frameworks cluster at 23-26%, while model rates range from 0.4% to 49.7%. Claude Opus 4.8 was vulnerable in 8 of 530 cells. A regex description scanner blocks 20 of 45 scenarios at 7.2% false positives, so metadata scanning is only a partial control.