Back to home
Blog
Security research, benchmark results, and insights from the MCPShield team.
researchsecuritydata
We Scanned 1,585 MCP Servers. 1 in 6 Had Hidden AI Manipulation.
The largest MCP security study to date. 18.4% fail with critical issues. 17.1% have semantic attacks invisible to regex. We built 53 adversarial servers and an LLM judge to catch what patterns can't.
Read more
researchadversarialbenchmarks
We Tested 4 Published Evasion Techniques Against Our MCP Scanner. All Failed.
We crafted 20 adversarial tool descriptions using techniques from the AutoMalTool paper and tested them against MCPShield. Both our regex engine and LLM judge achieved 100% detection.
Read more