ai-security-lab - 🤖 Advanced AI/LLM Security Testing Framework. Incl

项目说明

🤖 AI Security Lab

<di="center">

!
!License
!Python
!AI
!Status

Ad AI/LLM Security Testing & Red Teaming Framework

Features • Installation • Tools • Jailbreaks • Documentation

</div>

---

🌟 O

AI Security Lab is a comprehensi for testing and e the security of Large Language Models (LLMs) and AI systems. This toolkit includes cutting-edge jailbreak techniques, prompt injection methods, ad testing tools, and automated scanners specifically designed for AI/ML systems.

🎯 Mission

As AI systems become increasingly integrated into critical applications, understanding their is essential. This lab provides security researchers, red teamers, and AI de with the tools to identify and mitigate risks in LLM deployments.

🔬 What We Test

---

✨ Features

Core Capabilities

| Feature | Description | Status |
|---------|-------------|--------|
| Automated Jailbreaking | 50+ pro techniques | ✅ Active |
| Prompt Injection Scanner | Detect and exploit prompt | ✅ Active |
| LLM Scanner | Automated security assessment (Garak, PyRIT) | ✅ Active |
| Ad Testing | Generate ad examples | ✅ Active |
| Red Team Prompts | 1000+ curated attack prompts | ✅ Active |
| API Fuzzing | Test AI API endpoints | ✅ Active |
| Bias Detection | Identify discriminatory outputs | ✅ Active |
| Data Extraction | Attempt to extract training data | ✅ Active |
| Multi-Model Support | GPT-4, Claude, Gemini, LLaMA, etc. | ✅ Active |
| Reporting | Automated reports | ✅ Active |

---

🚀 Installation

Quick Start

Prerequisites

Full Installation

---

🛠️ Tools Included

🎯 Core Testing Tools

#### 1. Garak - LLM Scanner

What it tests:

#### 2. PyRIT - Python Risk Identification Toolkit

Capabilities:

#### 3. Promptfoo - LLM Testing Framework

Features:

🔓 Jailbreak Tools

#### 4. Jailbreak Tester (Custom)

Included Techniques:

#### 5. Prompt Injection Scanner

🎭 Ad Testing

#### 6. Ad Prompt Generator

#### 7. Token Smuggling Tool

📊 Analysis Tools

#### 8. Bias Detector

#### 9. Data Extraction Tester

---

🔓 Jailbreak Techniques

Classic Jailbreaks

#### DAN (Do Anything Now)

#### STAN (Strive To Avoid Norms)

#### De Mode

Ad Techniques

#### Multi-Turn Jailbreak

#### Role-Playing Jailbreak

#### Encoding-Based Bypass

🆕 Latest Techniques (2026)

See jailbreaks/ directory for complete collection.

---

📚 Prompt Library

Categories

1. Malicious Intent Detection (500+ prompts)
- Illegal acti
- Harmful content
- Misinformation
- Hate speech

2. Safety Filter Bypass (300+ prompts)
- Content policy
- Ethical boundary testing
- Alignment breaking

3. Data Extraction (200+ prompts)
- Training data extraction
- PII extraction
- API key leakage

4. Ad Examples (1000+ prompts)
- Misclassification triggers
- Confidence manipulation
- Output poisoning

---

🎓 Testing Methodology

Phase 1: Reconnaissance

Phase 2: Scanning

Phase 3: Jailbreak Testing

Phase 4: Ad Testing

Phase 5: Reporting

---

📊 Example Results

Jailbreak Success Rates (January 2026)

| Model | DAN | STAN | De Mode | Multi-Turn | O |
|-------|-----|------|----------------|------------|---------|
| GPT-4 | 12% | 8% | 5% | 45% | 18% |
| Claude 3 | 5% | 3% | 2% | 38% | 12% |
| Gemini Pro | 15% | 10% | 8% | 52% | 21% |
| LLaMA 3 | 35% | 28% | 25% | 68% | 39% |

*Note: Success rates vary based on prompt engineering and model updates*

Common Found

1. Prompt Injection - 85% of models
2. Context Manipulation - 72%
3. Role-Playing Exploits - 68%
4. Encoding Bypass - 45%
5. Multi-Turn Attacks - 91%

---

⚠️ Legal & Ethical Guidelines

✅ Authorized Use

❌ Prohibited Use

🛡️ Responsible Disclosure

If you disco:

1. Do not exploit the maliciously
2. Report to the through proper channels
3. Allow time for the to patch (typically 90 days)
4. Disclose responsibly after confirmation

Disclosure Contacts:

---

🔬 Research & Papers

Recommended Reading

1. "Uni and Transferable Ad Attacks on Aligned Language Models" (2023)
2. "Jailbroken: How Does LLM Safety Training Fail?" (2023)
3. "Exploiting Programmatic Behavior of LLMs" (2024)
4. "Red Teaming Language Models with Language Models" (2022)
5. "Prompt Injection Attacks and Defenses in LLM-Integrated Applications" (2024)

Conferences

---

🛠️ Configuration

API Keys Setup

Create .env file:

Custom Model Configuration

Edit config.yaml:

---

📖 Documentation

Detailed documentation a in docs/:

| Document | Description |
|----------|-------------|
| SETUP.md | Installation and configuration |
| JAILBREAKS.md | Complete jailbreak techniques |
| TOOLS.md | Tool usage guides |
| METHODOLOGY.md | Testing methodology |
| ETHICS.md | Ethical guidelines |
| API.md | API documentation |
| CONTRIBUTING.md | Contribution guidelines |

---

🤝 Contributing

Contributions welcome! We especially need:

See CONTRIBUTING.md for guidelines.

---

🔄 Updates

Latest Updates (v1.0.0 - January 2026)

Roadmap

v1.1 (Q2 2026)

v2.0 (Q4 2026)

---

📜 License

MIT License - see LICENSE for details.

Disclaimer: This toolkit is for authorized security testing and research only. Unauthorized use may laws and Terms of Service.

---

🙏 Acknowledgments

---

📞 Contact

---

⭐ Star History

If this project helps your research, please give it a star! ⭐

---

<di="center">

🤖 Securing AI, One Test at a Time 🤖

Use Responsibly. Test Ethically. Report .

⬆ Back to Top

</div>

项目信息

开发语言：未知

Star 数：11

技术要点

该项目属于安全软件方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕🤖 Advanced AI/LLM Security Testing Framework. Includes 50+ jailbreak techniques, prompt injection tools, and automated vulnerability scanners for GPT-4, Claude, Gemini, and more.展开，提供了可运行的实现，对理解相关安全机制、工具原理与攻防技术具有较高的研究价值。

从技术实现角度看，此类项目的核心在于对底层原理的深入理解与工程化落地。开发者在研究过程中可以结合官方文档与社区讨论，梳理项目的初始化流程、核心模块职责与数据流转，逐步掌握其中的技术细节与设计思路。

使用与学习建议

1. 下载源码压缩包，解压到本地环境。

2. 根据项目类型准备对应的运行环境与依赖。

3. 按照 README 中的步骤完成配置与部署。

4. 在隔离的测试环境中运行验证，避免影响生产系统。

免责声明

本资源来源于开源社区，仅供安全研究、技术学习与合法授权范围内的攻防演练使用，请勿用于任何非法或侵害他人权益的用途。使用本项目产生的一切后果由使用者自行承担。