One in Five Popular MCP Servers Harbors Security Flaws

A security scan of the top 100 MCP servers on Smithery reveals that 22 contain vulnerabilities, with tool description injection being the most prevalent issue. This finding demonstrates that agentic AI systems are exposed to real security risks in production, as agents can be manipulated through seemingly innocent tool descriptions or encoding patterns. The widespread vulnerabilities highlight the urgent need for security standards in the MCP ecosystem as AI agents become more autonomous.