CyberSunil/LLMVault

An intentionally vulnerable OWASP LLM Top 10 training platform for AI Security, Prompt Injection, RAG Security, Agent Security, and GenAI penetration testing.

What it solves

LLMVault provides a safe, hands-on environment for developers and security researchers to learn about the OWASP Top 10 for LLM Applications. It solves the problem of lacking practical, "capture-the-flag" (CTF) style training ranges where users can practice exploiting and then defending against common AI vulnerabilities without risking production systems.

How it works

The platform operates in two distinct modes:

  • Play Mode: A scored range featuring 25 labs across three tiers (Core, Advanced, and Expert). These labs use scripted assistants to ensure reproducible results, allowing users to practice specific techniques until they become muscle memory.
  • Live Mode: Connects to a real local LLM (via Ollama or OpenAI-compatible endpoints). In this mode, secrets are generated fresh per session, and the model's unpredictable nature requires users to genuinely manipulate the AI to leak information.

The system is built with Flask and can be deployed via Docker. It includes a dashboard to track progress, a scoring system for Play Mode, and integrated reference panels that explain the OWASP categories before each lab and the necessary fixes after completion.

Who it’s for

It is designed for security educators, developers, and AI practitioners who want to learn how to identify, exploit, and mitigate LLM-specific security risks like prompt injection, data poisoning, and excessive agency.

Highlights

  • Tiered Learning Path: Progresses from 10 core OWASP labs to 10 advanced multi-turn conversational challenges and 5 encrypted expert labs.
  • Real-Model Integration: Live Mode allows testing against actual local models (e.g., Qwen 2.5) to experience unpredictable AI behavior.
  • Comprehensive Coverage: Covers the full OWASP LLM Top 10, including multimodal indirect injection and downstream output handling.
  • Integrated Education: Pairs every attack scenario with a detailed explanation of the defense and the fix.
  • Gamified Experience: Features a scoring system, ranks (from Initiate to Vault Master), and shareable completion cards.

Related

  • Project
    CVHub520/X-AnyLabeling-Server

    X‑AnyLabeling‑Server is a lightweight, extensible Python server that wraps AI models as a REST API for the X‑AnyLabeling annotation tool. It separates service management from model code, offers plug‑in model integration, production‑grade logging, concurrency control, and flexible configuration, enabling fast auto‑labeling workflows.

  • Project
    wanyichen06/LLMInternSkill

    An evidence-bound resume toolkit for AI internship seekers that polishes resumes, tailors them to JDs, and generates interview prep based on verifiable technical evidence.

  • Project
    IBM/prompt-declaration-language

    A declarative YAML-based language for creating reliable, composable LLM prompts and integrating them into software systems with support for control flow and code execution.

  • Project
    JuliaDiff/DifferentiationInterface.jl

    A unified interface for various automatic differentiation backends in Julia, providing a consistent syntax for computing gradients, Jacobians, and Hessians.