From 9% to 63%: Letting a Benchmark Build a Better MCP Security Proxy

Short summary

A detailed account of building mcp-bastion, an MCP security proxy, iteratively improving its attack coverage from 9% to 63% using a custom benchmark (mcp-defense-bench). The benchmark uses 24 MCP attack vectors with matched benign controls, measuring only runtime behavior. Key improvements included scanning tool results, validating arguments, cross-tool correlation for split-payload attacks, and cross-server data-flow tracking.

• •Built mcp-defense-bench: a vendor-neutral benchmark with 24 MCP attack vectors and matched benign controls
• •Iteratively improved mcp-bastion proxy coverage from 9% to 63% by addressing measured gaps
• •First tool to catch ShareLock split-payload attacks via cross-tool correlation

Generated with AI, which can make mistakes.