RAG-GenAI-Secure

A portfolio project demonstrating DevSecOps and AI/GenAI security skills, built as a hands-on showcase for a senior security leadership role.

The project has two parts: building and attacking a RAG-based AI application (Part A), and running it through a formal application security program (Part B).

---

Part A — AI / GenAI Security

A retrieval-augmented generation (RAG) pipeline built on a fully self-hosted local stack, then deliberately attacked and hardened.

Stack

Component	Technology
LLM inference	Ollama (llama3, llama3.2:1b)
Embeddings	Ollama (nomic-embed-text)
Vector database	Qdrant (Docker, self-hosted)
RAG framework	LlamaIndex
Runtime	Python 3.12
Host	Fedora 44 ThinkPad X1 Carbon (son-of-anton)

Attack Scenarios Demonstrated

Attack 1 — Indirect Prompt Injection (OWASP LLM01)

• Planted a malicious instruction inside a vendor policy PDF
• Showed the instruction being retrieved and executed by the LLM, causing it to refuse legitimate user questions
• Fixed with a regex-based input sanitizer (src/sanitizer.py) that strips injection patterns from retrieved chunks before prompt construction

Attack 2 — Sensitive Data Leakage (OWASP LLM06)

• Planted fake database credentials, API keys, and PII in a runbook document
• Showed the LLM leaking the full PostgreSQL connection string and SSN verbatim in response to direct queries
• Fixed with a DLP output scanner (src/output_scanner.py) that detects and redacts sensitive patterns before the response reaches the user

Key Files

src/

ingest.py # PDF ingestion → chunking → embedding → Qdrant

query.py # Query → retrieval → augmented prompt → LLM response

sanitizer.py # Input sanitizer: strips prompt injection patterns

output_scanner.py # Output DLP scanner: redacts sensitive data patterns

scripts/

plant_injection.py # Attack 1: generates a poisoned PDF

plant_leakage.py # Attack 2: generates a document with planted secrets

data/source_docs/ # Acme Corp synthetic policy documents (corpus)

---

Part B — Application Security Program

The same application put through a formal enterprise AppSec gate.

Controls Implemented

Control	Tool	Status
Threat modeling	STRIDE	✅ Complete — 4 components, 10 threats
Static code analysis	Semgrep	✅ Automated in CI
Dependency scanning	pip-audit	✅ Automated in CI
CI/CD security pipeline	GitLab CI + GitHub Actions	✅ Live
Vulnerability tracking	VULN_TRACKER.md	✅ 18 findings tracked
Metrics reporting	METRICS_SUMMARY.md	✅ Leadership-ready

CI Pipelines

Both platforms run SAST (Semgrep) and SCA (pip-audit) automatically on every push to main.

• GitLab CI: self-hosted runner on son-of-anton (Docker executor)
• GitHub Actions: GitHub-hosted Ubuntu runner

Key Documents

docs/

THREAT_MODEL.md # STRIDE threat model — 10 findings across 4 components

VULN_TRACKER.md # Full vulnerability tracker — 18 findings, severity, status

SCAN_RESULTS.md # SAST and SCA scan results summary

METRICS_SUMMARY.md # Leadership metrics one-pager

.gitlab-ci.yml # GitLab CI pipeline

.github/workflows/

security-scans.yml # GitHub Actions workflow

---

Real Engineering Issues Encountered and Resolved

This project was built hands-on. Real issues hit along the way:

• SELinux blocking Qdrant volume mount — fixed with :Z Docker flag
• ROCm GPU detection failure — Intel iGPU on X1 Carbon; no AMD GPU; confirmed CPU-only inference path and documented as a known constraint
• llama3 CPU timeout — switched to llama3.2:1b for development; llama3 retained for final demo use
• SimpleDirectoryReader reading raw PDF bytes silently — fixed by installing llama-index-readers-file; root cause documented
• Python 3.14 numpy wheel unavailable — downgraded project venv to Python 3.12 where all ML ecosystem wheels are prebuilt

---

Running Locally

# Prerequisites: Docker, Ollama, Python 3.12

# Start Qdrant
docker run -d --name qdrant -p 6333:6333 \
  -v ~/qdrant_storage:/qdrant/storage:Z qdrant/qdrant

# Pull models
ollama pull llama3.2:1b
ollama pull nomic-embed-text

# Install dependencies
python3.12 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# Ingest documents
python src/ingest.py

# Query
python src/query.py "What is required of Tier 1 vendors?"

---

Frameworks Referenced

• OWASP LLM Top 10
• STRIDE Threat Modeling
• NIST AI RMF (Govern, Map, Measure, Manage)