A guard-agnostic benchmark for MCP injection / exfil / tool-poisoning detection. Versioned cases + a language-agnostic runner that scores any guard through its own published CLI.
- capability exposure inferred + 16
- recent drift inferred + 20
- tool safety inferred + 12
inferred
The A–E grade is our heuristic synthesis — a "review this" prompt, not a verdict. Each factor is tagged by what backs it: attested (a verifiable record), reported (a third party's claim), or inferred (our own heuristic, e.g. permissions). See methodology.
graded 6m ago · see ecosystem CVEs →
- C · 40 → C · 48
- A · 0 → C · 40
No known CVEs for this server.
- high dangerous code
committed secret: GitHub token, private key
analyzed commit a21ab71 · analyzer v33 · 3d ago
danger signals2
- committed secret GitHub token getmcpm-mcp-guardbench-a21ab71/cases/attacks/credential-egress-github-pat.json :18
ghp_A1…(40 chars, redacted) - committed secret private key getmcpm-mcp-guardbench-a21ab71/cases/attacks/credential-egress-private-key.json :18
PEM private key block (redacted)
- recent drift +20 capability drift →
[](https://mcpobservatory.com/servers/github:getmcpm/mcp-guardbench/security) Heuristic, inferred signals — false positives (legitimately powerful tools, forks, language ports) are expected. Treat each as "review this", not a verdict. See the ecosystem-wide picture on the security hub, or the fleet security of getmcpm.