AI Security Learning Lab

(LLMs · Secure RAG · Cloud · MLOps)

This README is available in:

• 🇬🇧 EN English (this document)
• 🇪🇸 ES Español → README_ES.md

---

Overview

This repository is my personal AI Security learning lab and project index, documenting my transition from Cloud Security into AI / LLM Security, with the long-term goal of an AI Cloud Security Architect profile.

The focus is on building real AI systems from scratch, making security and architectural decisions explicit, validating them through hands-on testing, and documenting both results and failures.

This is not a tutorial repository.
It is an engineering-first, security-first exploration of modern LLM-based systems.

---

Learning Objectives

• Understand how LLMs work internally (context windows, prompts, generation behavior).
• Design and build secure Retrieval-Augmented Generation (RAG) pipelines.
• Apply practical AI security controls, inspired by the OWASP LLM Top 10:
  • Prompt injection and jailbreaks
  • Data leakage and PII exposure
  • Ingestion risks and data poisoning
  • Abuse prevention and auditability
• Prepare systems for enterprise-like environments:
  • Multi-environment configuration
  • Observability and structured logging
  • Cloud-ready architecture
  • MLOps foundations

---

Projects

🔬 Secure RAG from Scratch — Version 1 (Legacy Lab)

📁 secure-rag-from-scratch/

This project represents the first exploratory iteration of Secure RAG.

Its goal was to:

• Understand the core mechanics of RAG systems
• Experiment with early input security controls
• Validate basic assumptions before introducing infrastructure complexity

Key characteristics of V1:

• Local-only execution
• In-memory vector store
• Mock LLM (provider-independent)
• Initial prompt injection detection
• Manual testing and documentation

⚠️ This version is now considered a legacy learning lab.

👉 The actively developed and extended version is Secure RAG v2, available as a standalone repository.

---

🚀 Secure RAG from Scratch — Version 2 (Main Project)

🔗 Repository:
https://github.com/RescribanoSecurity/secure-rag-from-scratch-v2

Secure RAG v2 is the main, production-grade learning project, evolving directly from the lessons learned in V1.

Focus areas include:

• Modular RAG pipeline architecture
• Input and output security controls
• PII detection, redaction, and output blocking
• OWASP LLM Top 10 mapping
• Dockerized infrastructure
• Real vector database (Qdrant)
• Auditability and request tracing
• Visual validation via Streamlit UI

This project is:

• Runnable
• Testable
• Evidence-driven
• Explicit about what is implemented and what is not

📄 Full documentation, screenshots, and technical presentations are maintained inside the V2 repository.

---

Design Philosophy

• Incremental by phases: each phase is stable, reviewable, and extensible.
• Security as a first-class concern, not an afterthought.
• Decoupled architecture: security controls are independent of LLM providers or vector stores.
• Local-first, cloud-ready: reduce early complexity while designing for future scale.
• Failures are documented, not hidden.

---

Roadmap (High-Level)

• Secure RAG local baseline (V1)
• Input security controls
• Output security (PII detection, redaction, blocking)
• Manual validation with documented evidence
• Cloud-native vector stores (OpenSearch / Azure AI Search)
• Authentication and identity-aware logging
• Persistent audit logs
• CI/CD and automated security testing
• Threat modeling and full OWASP LLM Top 10 enforcement

---

Why this repository exists

Modern AI systems will be attacked.

Understanding how to:

• design them securely,
• validate security controls,
• detect abuse,
• audit behavior,
• and evolve architectures safely,

is becoming a core security skill.

This repository documents that journey.