Scan your MCP server — $49

Upload nothing. Give us a public repo URL and we scan it statically (no code execution), then email you a findings report with file/line locations and reproduction notes.

Public repositories only (github.com / gitlab.com / gitee.com). Delivery within 24 hours. Payment: USDT (TRC20). Static analysis only — no code is executed and no source is kept.

Not ready to pay? Get a free scan first — top 5 findings emailed, no card, no account.

Need it wired into CI? $299 deep review — findings grouped with fix actions, SARIF export, a ready workflow, and one re-scan within 30 days.

mcp-audit — Sample Findings Report

This is the exact shape of the report you receive for your repository.
It was generated from tests/sample_vuln.py, an intentionally-vulnerable fixture
shipped with the scanner — not a real client repo. Sample is based on public code (the project's own test fixture), already sanitized; no client data is shown.

Target: tests/sample_vuln.py (fixture)

Engine: mcp-audit v1 (pure stdlib, MIT)

Scan date: 2026-09-18


Severity summary

SeverityCount
CRITICAL3
HIGH4
MEDIUM0
LOW1
Total8

Findings

CRITICAL

1. call:os.system — shell 命令执行

  • tests/sample_vuln.py:10
  • The tool calls os.system(...). Any input that reaches this call becomes a shell command.
  • 2. shell_injection — subprocess shell=True

  • tests/sample_vuln.py:12
  • subprocess invoked with shell=True is a command-injection surface: unquoted user input is interpreted by the shell.
  • 3. call:eval — 动态代码执行

  • tests/sample_vuln.py:14
  • eval() executes arbitrary code from its argument. If that argument can be influenced by input, the tool is remote-code-executable.
  • HIGH

    4. import:subprocess — 导入 subprocess

  • tests/sample_vuln.py:4
  • Imports subprocess, which can execute shell commands (see CRITICAL #1/#2).
  • 5. call:subprocess.run — 子进程调用

  • tests/sample_vuln.py:12
  • A subprocess is spawned; confirm whether shell=True is set (see CRITICAL #2).
  • 6. secret_env — 读取敏感环境变量 AWS_SECRET_ACCESS_KEY

  • tests/sample_vuln.py:20
  • Reads AWS_SECRET_ACCESS_KEY from the environment. A malicious tool can exfiltrate cloud credentials.
  • 7. secret_env — 读取敏感环境变量 PRIVATE_KEY

  • tests/sample_vuln.py:21
  • Reads PRIVATE_KEY from the environment. A malicious tool can exfiltrate wallet/signing keys.
  • LOW

    8. import:os — 导入 os

  • tests/sample_vuln.py:3
  • Imports os; os.system/os.popen carry shell risk (see CRITICAL #1).

  • Reproduction notes (CRITICAL/HIGH)

  • CRITICAL #1 (os.system) — if any tool parameter (e.g. a filename, a URL, a query string) is concatenated into the os.system string, a malicious MCP client can inject ; curl ... | sh and run arbitrary commands on the host running the server.
  • CRITICAL #3 (eval) — if eval() receives any remotely-influenced string, the server is effectively arbitrary-code-executable. This is the single highest-severity pattern an MCP server can contain.
  • HIGH #6/#7 (secret_env) — any tool that reads AWS_SECRET_ACCESS_KEY / PRIVATE_KEY from the environment can forward those values to an attacker-controlled endpoint. Even a "read-only" tool description can hide this.

  • *Static analysis flags patterns, not proven exploits. Each finding above needs a manual read to confirm whether untrusted input can actually reach the sink — manual review of individual findings is available on request, quoted per engagement (reply to the delivery email with the order id).*

    Run it yourself, free: the scanner is MIT, stdlib-only, one ~8KB file - gitlab.com/jamKing168/mcp-audit
    Want this run on your repo?
    $49 self-serve: one public repo, scanned statically, findings emailed with file/line locations - order from the form at the top of this page.
    Deeper engagements (CI wiring, per-finding manual confirmation) are quoted per engagement: SCOPE.md.
    To start, DM @king94168 (or reply to the email you received).