LLM Prompt Injection Security Lab

Research and defense implementation for prompt injection vulnerabilities in LLM applications.
Hands-on lab work completed using SecureFlag; defenses implemented with the llm_guard toolkit.

---

Overview

Completed SecureFlag professional security lab demonstrating prompt injection vulnerabilities and defenses for LLM applications. This repo documents attack reproduction, mitigation design, and verification for the Prompt Injection — Leaks Sensitive Content lab (and provides a pattern you can reuse across other OWASP LLM Top 10 labs).

Project Highlights

• Platform: SecureFlag professional security training
• Completion Date: October 17, 2025
• Reproduced system-prompt exfiltration via a reflection/structured-output attack
• Implemented a model-based defense using LLM Guard and added output sanitization
• Delivered a test suite, write-up, and reusable prompt-security module

What I Learned

• How to identify and reproduce prompt injection attacks that lead to information exfiltration.
• Attack vectors: direct override, role manipulation, context switching, and reflection.
• How to integrate ML-based defenses (llm_guard) as a pre-LLM gate and programmatic output sanitizers.
• Practice in secure-by-design LLM handling and defense-in-depth.

Repository Contents

• /vulnerability-research - Attack patterns and exploitation techniques
• /defense-implementation - Security controls using LLM Guard
• /documentation - Detailed write-ups and analysis
• /test-cases - Sample prompts for testing defenses

Key Achievements

✅ Successfully extracted system prompts from vulnerable LLM
✅ Implemented prompt injection detection with LLM Guard
✅ Created comprehensive documentation of attack patterns
✅ Developed reusable security module

Technologies Used

• Python 3.x
• LLM Guard Security Toolkit
• OWASP Security Methodology

Getting Started

pip install -r requirements.txt
python test-cases/injection_tests.py