MCP Tool Poisoning Hits Every Framework

🔥 What happened Researchers Farooq Shaikh and Biagio Montaruli ran the first cross-framework measurement of MCP tool poisoning with real tool execution. Their tool mcpvenom tested 12 frameworks, 18 models, and 45 attack scenarios – 22.9% of cells were vulnerable. 💡 Why it matters Not a single one of the twelve frameworks implements an intentional defense – all cluster between 23% and 26% attack success. Vulnerability is model-level, not framework-level: Claude Opus 4.8 was vulnerable in just 8 of 530 cells, while other models hit 49.7%. ⚡ Our take Running MCP tools without sandboxing is Russian roulette with your secrets. Metadata scanning is no shield – it only catches 20 of 45 scenarios.