Skip to content

README_SPACE.md

Latest commit

 

History

History
191 lines (149 loc) · 8.23 KB

README_SPACE.md

File metadata and controls

191 lines (149 loc) · 8.23 KB
Raw
title AI Security Education Interactive Demo
emoji 🛡️
colorFrom red
colorTo blue
sdk gradio
sdk_version 4.44.0
app_file app.py
pinned false
license apache-2.0
models
Zen0/Vulnerable-Edu-Qwen3B
tags
security
education
jailbreak
defence
australian-compliance

🎓 AI Security Education - Interactive Demo

Live demonstration of AI jailbreak attacks and defence systems

🚀 Try It Now!

This Space lets you:

  • 🔴 Attack a vulnerable AI model - See jailbreaks work in real-time
  • 🛡️ Test defence systems - Watch attacks get blocked
  • ⚖️ Compare side-by-side - Vulnerable vs protected models
  • 🇦🇺 Learn Australian compliance - Privacy Act 1988 context

🎯 What You'll Learn

Jailbreak Techniques

  • DAN (Do Anything Now) - Classic instruction override
  • Skeleton Key - Microsoft's 2024 discovery
  • Base64 Encoding - Obfuscation attacks
  • Role Playing - Persona jailbreaks
  • System Extraction - Prompt leaking

Defence Architecture

  • 7-Layer Defence System
    1. Input Validation
    2. Prompt Sanitisation
    3. Context Isolation
    4. Output Filtering
    5. Monitoring & Logging
    6. Rate Limiting
    7. Human Oversight

Australian Compliance

  • Privacy Act 1988 — post-2024 reform (tiered penalties up to $50M / 3× benefit / 30% turnover for serious/repeated breaches; new statutory tort for serious invasions of privacy commenced 10 Jun 2025)
  • ACSC Essential Eight
  • Notifiable Data Breaches scheme + OAIC 30-day notification
  • APRA CPS 234, PSPF / ISM context

📚 Full Course

This Space is part of a complete AI Security Education course:

Repository: Benjamin-KY/AISecurityModel

Includes:

  • 🎓 18 progressive Jupyter notebooks across three tracks
    • 🟢 Foundational (nb01–nb06): introduction → defence-in-depth at the prompt boundary
    • 🟠 Advanced (nb07–nb15): red-teaming, monitoring, fine-tuning, multi-modal, supply chain, incident response
    • 🟣 2026 Architectural Capstone (nb16–nb18): agent/MCP tool misuse, RAG-layer prompt injection, harness paradigm
  • 💻 Hundreds of small, executable, individually-runnable code cells (refactored to gold-standard pedagogy in v2.2.x)
  • 📖 Comprehensive educator guide with five course formats (2 h workshop → 5-day intensive)
  • 🧭 Strategic positioning doc and operational reading order across the three-repo constellation
  • 🔬 XAI & interpretability tools
  • 🛡️ Defence code patterns (not production-ready as shipped — see disclaimer)
  • 🇦🇺 Australian regulatory compliance grounding

Perfect for:

  • University AI security courses
  • Security professional training
  • Australian organisations deploying AI
  • Researchers studying LLM vulnerabilities

🌐 Part of a three-repo constellation

This Space is the attacks lab half of the picture. The whole story spans three repositories:

Repo Role
AISecurityModel (this Space) Model & prompt-layer attacks course + bridge into the architectural layer
harmless-harnesses Architectural / harness-layer course — 29 modules across F/C/P/T/CAP tracks
sa-sovereign-llm-harness Research codebase + SA-GOV-BENCH evaluation methodology (May 2026 results)

Recommended path: AISecurityModel → harmless-harnesses → sa-sovereign-llm-harness. See docs/POSITIONING.md for the full map.

🔬 Educational Pattern

Vulnerable-Then-Educate:

  1. Model shows the vulnerability (complies with jailbreak)
  2. Provides educational analysis
  3. Explains prevention strategies
  4. References compliance requirements

⚠️ This model is INTENTIONALLY VULNERABLE for education

🛠️ Technical Details

Model: Qwen2.5-3B fine-tuned with LoRA Parameters: 3 billion Size: ~6 GB (FP16) Training: 15 vulnerability examples Hardware: Optimised for RTX 3060 12GB

📖 How to Use This Space

  1. Choose a Tab:

    • 🔴 Vulnerable Model - See attacks work
    • 🛡️ Defended Model - See defences block attacks
    • ⚖️ Comparison - See both side-by-side

  2. Select an Attack:

    • Use dropdown for pre-made examples
    • Or type your own custom attack

  3. Click the Button:

    • Watch the response in real-time
    • Read the educational analysis
    • Understand the security implications

  4. Learn & Experiment:

    • Try different attack types
    • Modify existing attacks
    • See what gets blocked and why

🇦🇺 Australian Context

All educational content includes:

  • Privacy Act 1988 references
  • ACSC Essential Eight controls
  • Notifiable Data Breaches scheme
  • Australian English orthography
  • Local PII patterns (TFN, Medicare, etc.)

🎓 Related Resources

⚠️ Important Disclaimers

  1. Educational Use Only — This model is intentionally vulnerable
  2. Not for Production — Use defence examples for real deployments; for the architectural layer above the model, see the harmless-harnesses sibling course
  3. Necessary, not sufficient — Defending only at the prompt boundary (what this Space demonstrates) does not eliminate harm. The 2026 Architectural Capstone notebooks (nb16–nb18) and harmless-harnesses make the case explicitly
  4. Supervised Use — For educational and research contexts
  5. Ethical Use — Do not use techniques maliciously
  6. IDSov boundary — This course sits inside the dominant Western technical-governance frame and does not model Indigenous Data Sovereignty practice as a first-class governance authority. For the full positioning, see docs/the-harness-paradigm.md in sa-sovereign-llm-harness (referenced verbatim in harmless-harnesses F0 §6 and F2 §6)

📜 Citation

If you use this in research or education:

@software{aisecurityedu2026,
  author = {Kereopa-Yorke, Benjamin},
  title = {AISecurityModel: AI Security Education Course},
  year = {2026},
  url = {https://github.com/Benjamin-KY/AISecurityModel},
  note = {Interactive demo: https://huggingface.co/spaces/Zen0/AI-Security-Education}
}

🤝 Contributing

Found an issue? Have suggestions?

  • Open an issue on GitHub
  • Submit a PR with improvements

📧 Contact

Author: Benjamin-KY GitHub: Benjamin-KY Model: Zen0/Vulnerable-Edu-Qwen3B

Built with ❤️ for AI Security Education by Ben Kereopa-Yorke 🇦🇺 Australian Privacy Act 1988 context — post-2024 reform aware