The Problem
Companies adopting AI agents and MCP (Model Context Protocol) servers have almost no way to check them for security holes before deploying them. A single unsafe tool declaration, an over-permissioned agent, or a leaked secret in a config file can open the door to command injection, tool poisoning, or data exfiltration -- and most teams only find out after an incident.
What I Built
mcpscan: a supply-chain security scanner purpose-built for MCP servers and AI agent deployments. It runs 22 static analysis rules mapped directly to the OWASP MCP Top 10:2025, covering command injection, tool poisoning, secret leakage, and CI/CD risk in GitHub Actions.
Proof It Works
I benchmarked the scanner against four real, disclosed 2026 CVEs affecting MCP infrastructure -- including a CVSS 9.1 remote code execution vulnerability -- and it correctly flagged all of them.
What This Means For You
If you are shipping AI agents, MCP servers, or Claude/GPT-based tooling into production, I can run this scan plus a deeper manual review against your stack before it goes live, or wire it in as a recurring gate in your CI pipeline. Available as a one-time security audit or an ongoing automated check.
0 likes · 3 views · Sep 16, 2026
Open-source MCP security scanner that caught 4 real 2026 CVEs, including a CVSS 9.1 RCE, before a single exploit ran.