Open-source MCP security proxy improves from 9% to 63% attack coverage via benchmark-driven development

An open-source security proxy for AI agents using the Model Context Protocol (MCP) improved its attack coverage from 9% to 63% by iterating against a custom benchmark. The benchmark includes 24 attack vectors with matched benign controls to avoid false positives. The proxy, mcp-bastion, now blocks or detects most known MCP attacks, demonstrating a data-driven approach to AI agent security.