IVR MCP

IVR-MCP

基于模型上下文协议（MCP）的交互式语音应答（IVR）系统的人工智能渗透测试框架。AI代理协调5台MCP服务器，自主发现、映射和测试IVR系统的安全漏洞，包括CallerID/ANI/P-Asserted-Identity欺骗、DTMF注入、身份验证绕过和语音生物识别攻击。

核心方法是 迭代探测 --人工智能呼叫目标，记录呼叫，挂断电话，下载录音，用时间戳转录，分析提示和计时，然后用从转录中计时的DTMF输入再次呼叫。每次呼叫都会深入IVR树的一层。

版本

1.3.0

作者

加兰德·格莱斯纳 — gglessner@gmail.com

许可证

GNU通用公共许可证v3.0--请参阅 许可证

运作原理

┌─────────────────────────────────────────────────────────────────┐
│                    ITERATIVE IVR PROBING                        │
│                                                                 │
│  Probe 1: Call → Record → Hang up → Download → Transcribe       │
│           "Welcome. Press 1 for English..." [0.0s-4.5s]         │
│                                                                 │
│  Probe 2: Call → Record → DTMF 1 at 5.5s → Hang up → Transcr.   │
│           "For billing press 1, support press 2..." [5.8s-11s]  │
│                                                                 │
│  Probe 3: Call → Record → DTMF 1 at 5.5s → DTMF 2 at 12s → …    │
│           Third-level menu discovered...                        │
│                                                                 │
│  Each probe builds on the timestamps from the previous one.     │
└─────────────────────────────────────────────────────────────────┘

建筑

                    AI Agent (Cursor / MCP Client)
                   /     |        |        \        \
          Asterisk-MCP  ElevenLabs Deepgram Audio  VoiceFuzz
              |          -MCP       -MCP    -MCP    -MCP
              |           |          |       |        |
          Asterisk PBX  ElevenLabs Deepgram ffmpeg  ffmpeg
              |          API        API    /pydub   /pydub
          SIP Trunk                /
              \                   /
               Target IVR System
                      |
          data/ <-- engagement logs + artifacts

MCP服务器

所有MCP服务器都位于 MCPs/ 目录。

电话控制

语音和音频

转录

IVR最重要的方法

先决条件

• Python 3.10+
• FFmpeg 关于PATH（适用于音频MCP和VoiceFuzz MCP）
• 授权书 适用于所有测试目标
• 启用ARI/AMI和SIP中继的Asterisk PBX（v18+）
• ElevenLabs API密钥（用于语音合成）
• Deepcram API密钥（用于转录）

快速开始

1.克隆并安装

git clone https://github.com/gglessner/IVR-MCP.git
cd IVR-MCP

# Install all MCP servers
pip install -r MCPs/asterisk-mcp/requirements.txt
pip install -r MCPs/deepgram-mcp/requirements.txt
pip install -r MCPs/elevenlabs-mcp/requirements.txt
pip install -r MCPs/audio-mcp/requirements.txt
pip install -r MCPs/voicefuzz-mcp/requirements.txt

2.配置API密钥

光标 --编辑 .cursor/mcp.json 直接使用您的API密钥和凭据。

VS Code --无需手动编辑。 .vscode/mcp.json 用途 ${input:...} 变量——VS Code将在首次启动时提示您输入每个秘密。

两者都需要凭据：

• ARI_USERNAME / ARI_PASSWORD --Asterisk ARI证书
• AMI_USERNAME / AMI_PASSWORD --Asterisk AMI凭据
• SSH_HOST / SSH_USER / SSH_KEY --SSH访问Asterisk服务器
• DEEPGRAM_API_KEY -Deegram API密钥
• ELEVENLABS_API_KEY -ElevenLabs API密钥

3.在IDE中打开

打开 IVR-MCP 目录作为您的项目。两个IDE都是预先配置的：

光标：

• .cursor/mcp.json --自动注册所有5台MCP服务器
• .cursor/skills/ivr-pentest/SKILL.md --从加载完整技能的存根 skills/

VS代码（副本）：

• .vscode/mcp.json --使用安全输入提示自动注册所有5台MCP服务器
• skills/ivr-pentest/ --Copilot代理可以访问的全技能文档

无需手动启动服务器。这两个IDE都按需启动MCP服务器。看 文档/MCP_SETUP.md 用于高级配置。

入门指南——示例提示

配置后，只需告诉AI该做什么。以下是提示：

设置和连接

Connect to Asterisk and verify the connection is working.

List available ElevenLabs voices.

启动IVR互动

Start a new IVR pentest engagement against +18005551234. Use +14155551234 as my ANI.
Don't brute force anything, just map out what is available.

Read skills/ivr-pentest/SKILL.md and begin an IVR mapping engagement
for the target at +18005551234. Spoof my caller ID as +14155551234.

侦察与测绘

Make 3 baseline calls to the target and transcribe each one.

Explore DTMF option 2 from the main menu.

Work your way through all menu items 1 through 9 and map the full IVR tree.

认证测试

Here are the test credentials — ANI: +14155551234, last 4 SSN: 1234, PIN: 5678.
Log in and map the authenticated menu.

Test the ANI spoofing bypass — does the IVR trust CallerID for authentication?

语音和音频

Generate audio using the "Rachel" voice saying "I need to verify my account"
and save it to the engagement artifacts.

Take the baseline recording and fuzz it with VoiceFuzz fractal mode at 3 stages.
Then transcribe the fuzzed version to verify it's still intelligible.

分析和报告

Create an analysis of the current engagement.

Update the engagement log and analysis with your latest findings.

高级

Test the account takeover chain — can someone change the PIN, change the address,
and order a new card in a single call without re-authenticating?

Run a CallerID/PAI isolation test — make 4 calls with different header combinations
to determine which header the IVR trusts for identity.

项目结构

IVR-MCP/
├── .cursor/
│   ├── mcp.json                       # MCP server config (Cursor)
│   └── skills/ivr-pentest/
│       └── SKILL.md                   # Stub → loads full skill from skills/
├── .vscode/
│   └── mcp.json                       # MCP server config (VS Code) with secure input prompts
├── skills/
│   └── ivr-pentest/                   # Shared skill files (both IDEs)
│       ├── SKILL.md                   # Core skill index, engagement workflow
│       ├── asterisk-mcp.md            # Asterisk tools, iterative probing, ANI/PAI spoofing
│       ├── elevenlabs-mcp.md          # Voice synthesis tools
│       ├── audio-voicefuzz.md         # Audio processing + VoiceFuzz biometric fuzzing
│       ├── methodology.md            # Testing phases 1-7
│       └── vulnerabilities.md         # Vuln classification (IVR-001 to IVR-025), test cases
├── MCPs/
│   ├── asterisk-mcp/                  # Asterisk MCP server (ARI + AMI)
│   ├── elevenlabs-mcp/                # ElevenLabs voice synthesis MCP server
│   ├── deepgram-mcp/                  # Deepgram STT MCP server
│   ├── audio-mcp/                     # Audio processing MCP (pydub/ffmpeg)
│   └── voicefuzz-mcp/                 # Voiceprint biometric fuzzing MCP
├── data/                              # Engagement data (gitignored)
│   └── YYYYMMDD_HHMMSS_target/        #   Per-engagement directory
│       ├── log.md                     #     Timestamped engagement log
│       └── artifacts/                 #     Recordings, TTS audio, fuzzed audio
├── LICENSE                            # GNU GPL v3.0
└── README.md

技能参考

AI技能文档存在于 skills/ivr-pentest/ （在两个IDE中共享），并拆分为重点文件：

游标通过存根自动加载这些文件 .cursor/skills/ivr-pentest/SKILL.md.VS Code Copilot代理可以参考 skills/ 直接目录。

法律声明

所有IVR渗透测试必须在系统所有者的明确书面授权下进行。 未经授权测试IVR系统是非法的，可能违反联邦电信法，包括《计算机欺诈和滥用法案》（CFAA）和窃听法规。根据《呼叫者身份真实法案》，呼叫者ID欺骗被允许用于授权的安全测试，但如果是为了欺诈而进行的，则是非法的。