🎯 Goal
Understand the unique security risks of machine learning systems, from training data poisoning to adversarial inputs, model theft, and prompt injection in Large Language Models (LLMs). Learn to assess, attack, and defend AI pipelines in development and production.
🔑 Key Skills to Learn
- ML Basics: Supervised/unsupervised learning, neural networks, loss functions, training vs. inference, key model types (CNN, RNN, transformers).
- Adversarial Machine Learning: Evasion attacks (adversarial examples), poisoning attacks (data, model), model inversion, membership inference.
- LLM Security: Prompt injection (direct/indirect), jailbreaking, system prompt extraction, RAG poisoning, tool misuse, excessive agency.
- Supply Chain Risks in AI: Compromised pre-trained models (Hugging Face, PyTorch Hub), malicious pickle files, dependency confusion.
- AI Governance & OWASP Top 10 for LLMs/ML: Frameworks for secure AI development and deployment.
- ML Pipeline Security: Securing data ingestion, model training (GPU environments), model serving APIs, and monitoring.
- Defensive Techniques: Adversarial training, differential privacy, model watermarking, guardrails, input sanitization, rate limiting.
🛠️ Essential Tools & Technologies
| Tool | Purpose |
|---|---|
| Adversarial Robustness Toolbox (ART) | Library for evaluating and defending ML models against attacks. |
| CleverHans | Adversarial example library for TensorFlow, PyTorch, JAX. |
| Garak | LLM vulnerability scanner (prompt injection, data leakage, hallucination). |
| TextAttack / OpenAttack | Frameworks for generating adversarial text examples. |
| Fawkes | Image cloaking tool to protect facial recognition from inference. |
| ModelScan | Scans serialized ML models (pickle, H5, etc.) for malicious code. |
| Guardrails AI | Enforce structure on LLM outputs, detect jailbreaks. |
| Rebuff / Prompt Injection Detector | Open-source detection for injection attacks. |
📖 Free Learning Resources
- OWASP Top 10 for LLM Applications – Critical risks and mitigations for LLMs. Link
- OWASP Top 10 for Machine Learning – Foundational risks in traditional ML systems. Link
- MITRE ATLAS – Adversarial threat landscape framework, similar to ATT&CK. Link
- NIST AI 100-2 E2023 – Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations. Link
- Google’s Secure AI Framework (SAIF) – Concepts and controls for secure AI systems. Link
- NVIDIA AI Red Team Resources – Public blog posts on LLM security and red teaming methodologies. Link
- HiddenLayer’s Security Research Blog – Real-world AI attacks and defenses. Link
- Practical LLM Jailbreaking (YouTube talks) – DEF CON AI Village, Black Hat recordings. Search for “LLM jailbreak” and “AI Village”.
🔗 Roadmap Placement
- Prerequisites: Python scripting, basic ML concepts (can be learned alongside), and application security fundamentals (APIs, injection attacks).
- Directly Enables: AI/ML Security Engineer, AI Red Teamer, LLM Application Security, or roles focused on securing AI-enabled products.
🌐 Website Concept & Interactive Features
Your website is a roadmap visualizer for cybersecurity learning. The heart is an interactive skill tree that maps exactly the modules we built, from Level 0 (IT Fundamentals) to Level 4 (specializations like AI/ML Security, Telecom, GRC, etc.).
1. The Skill Tree (Roadmap Page)
Visual layout:
A top-to-bottom tree graph.
- Level 0 at the top → Level 1 below → Level 2 → Level 3 → Level 4 branches at the bottom.
- Nodes are colored by domain (Offensive – red, Defensive – blue, Cloud – purple, Telecom – orange, etc.).
- Edges (lines) connect prerequisites to advanced modules.
Interaction:
- Hover: Show a tooltip with the module’s short description and difficulty.
- Click: Open a sidebar or modal with full details (goal, key skills, tools, free resources) – or navigate to a dedicated module page.
- Zoom/Pan: Allow users to explore the tree freely.
- Progress tracking: Users can mark modules as “done” (saved to browser local storage). Completed nodes get a checkmark or different color, giving a sense of progression like a skill tree in a game.
2. Module Pages
Each module (e.g., /linux, /cloud-security) displays exactly the content we defined:
- Goal & overview
- Key skills (bullet list)
- Tools & technologies (table with purpose)
- Free resources (links)
- Roadmap placement (prerequisites, next steps)
These pages are generated from Markdown files—easy to maintain and update without touching code.
3. Tools & Resources Library
A searchable, filterable table of every tool and resource across all modules.
Filters by: domain (Offensive/Defensive/Cloud…), type (tool vs. learning resource), difficulty level.
Powered by a lightweight client‑side search library (like Fuse.js or Pagefind) for instant results.
4. Tech Stack & Hosting (Why)
| Purpose | Technology | Reason |
|---|---|---|
| Overall site | Astro (static site) | Perfect for content‑heavy sites; ships fast HTML with minimal JavaScript. |
| Interactive graph | Cytoscape.js (via a React “island”) | Purpose‑built for beautiful, interactive node‑edge graphs. Loads only on the roadmap page. |
| Styling | Tailwind CSS | Rapid, consistent design with a utility‑first approach. |
| Content | Markdown / MDX | Easy writing and maintenance of all module pages. |
| Hosting | Cloudflare Pages | Free, fast, auto‑deploy from GitHub. Seamless integration with Cloudflare Workers for future backend needs. |
5. Implementation Plan (High‑Level)
- Set up Astro project with Tailwind and React integration.
- Build the roadmap component – define all nodes/edges based on the module tree we created.
- Create Markdown files for each module using the exact structure we've established.
- Build module page template – an Astro page that reads the Markdown and renders it with navigation to prerequisites and next steps.