mcp-security-audit

Security auditor for MCP servers. Connects to any MCP server, enumerates its tools/resources/prompts, scans for injection patterns, classifies risk levels, and produces a scored report (0-100, grades A-F).

Works as a CLI tool (mcp-audit) or as an MCP server itself (mcp-security-audit).

Why it matters: there are now 16,000+ MCP servers in the public ecosystem — and almost none have been security-audited. Each one runs with tool access to a shell, filesystem, or network, and a single injection-laden tool description can hijack the agent that loads it. This finds the holes before an attacker does.

Want a professional audit report? We found 20 vulnerabilities in Microsoft's MCP servers. Get the same analysis for your server — $29 single audit.

Install

pip install mcp-security-audit

Related MCP server: meok-mcp-injection-scan-mcp

CLI Usage

Text report

mcp-audit scan --server "python -m my_mcp_server"

JSON report

mcp-audit scan-json --server "python -m my_mcp_server" --output report.json

With live injection tests

mcp-audit scan --server "python -m my_mcp_server" --live-tests

CI/CD

Exit code 0 for grade A/B, 1 for C/D/F:

mcp-audit scan --server "uvx some-server" || echo "Security audit failed"

MCP Server Mode

Use as an MCP server so AI assistants can audit other servers:

Claude Code

claude mcp add mcp-security-audit -- uvx mcp-security-audit

Manual

{
  "mcpServers": {
    "mcp-security-audit": {
      "command": "uvx",
      "args": ["mcp-security-audit"]
    }
  }
}

Tools exposed

What It Checks

1. Tool risk classification — categorizes every tool as SHELL / FILE / DATABASE / NETWORK / SAFE

2. Injection pattern scanning — scans tool and prompt descriptions for 75 injection patterns across 9 categories

3. Resource URI analysis — flags sensitive paths (.env, .ssh, credentials, etc.)

4. High-risk ratio — warns if >50% of tools are FILE or SHELL level

5. Undocumented tools — flags tools missing descriptions

6. Attack surface — warns on 20+ tools (large) or 50+ (very large)

7. Live injection tests (opt-in) — sends payloads to string-parameter tools

Scoring

Starts at 100, deducts per finding:

Bonus: +1 per documented tool (max +5).

Grades: A (90+), B (75+), C (60+), D (40+), F (<40)

Example Output

============================================================
  MCP SECURITY AUDIT REPORT
============================================================

  Server:  python -m agent_safety_mcp.server
  Tools:   13
  Score:   92/100  (Grade A)

------------------------------------------------------------
  TOOL CLASSIFICATION
------------------------------------------------------------
  Tool                           Risk       Matched
  cost_guard_configure           SAFE       -
  cost_guard_status              SAFE       -
  injection_scan                 SAFE       -
  trace_save                     FILE       !!save, file
  ...

Dependencies

• mcp — MCP protocol SDK

• ai-injection-guard — 75 prompt injection detection patterns across 9 categories

Part of the LuciferForge AI Agent Infrastructure Stack.

License

MIT