AIHEM: Learn AI Security By Breaking Things
Learn AI security hands-on with AIHEM: 33 challenges covering the OWASP LLM Top 10 and advanced AI attacks. The WebGoat for the AI era.
· Updated · 9 min read · Guard0 Team

When I was learning web security, I learned by breaking things.
OWASP's WebGoat taught me SQL injection. DVWA showed me XSS. Hack The Box turned concepts into muscle memory. The best way to understand how attacks work is to execute them yourself, in a safe environment.
But when I started exploring AI security, I found a gap. Where's the WebGoat for LLMs? The DVWA for AI agents? The safe environment where security professionals can learn prompt injection, tool abuse, and memory poisoning by actually exploiting them?
It didn't exist. So we built it.
Welcome to AIHEM, the AI Hacking Educational Module.
What is AIHEM?
AIHEM is an intentionally vulnerable AI application platform designed for security education. It's a sandbox where developers, security professionals, and researchers can learn AI security through hands-on exploitation.
The Numbers:
- 33 challenges across all difficulty levels
- 3,825 total points available
- Covers all ten OWASP LLM Top 10 categories
- Includes six advanced AI attack challenges
- Runs locally via Docker Compose
Think of it as a CTF platform specifically for AI security, with progressive challenges that teach real-world attack techniques.
Why We Built This
The AI Security Skills Gap
AI is everywhere, but AI security expertise is rare. Security teams are being asked to assess AI systems without having practiced the attacks. Developers are building agents without understanding how they can be exploited.
You can read about prompt injection in a paper. But do you really understand it until you've crafted a payload that makes an agent leak its system prompt? Until you've chained a multi-step attack that exfiltrates data through tool abuse?
Learning By Doing
The most effective security training is experiential:
- Understand the concept
- Attempt the exploit
- Fail and iterate
- Succeed and understand why
- Apply to real-world defenses
AIHEM provides the environment for steps 2 to 4. Each challenge teaches a specific technique, with difficulty progression from beginner to expert.
Safe Harbor
You can't practice prompt injection on production systems (please don't). You need a sandbox where:
- It's legal and ethical to attack
- You can try aggressive techniques without consequences
- Failure doesn't affect real data
- Success provides clear feedback
AIHEM is that sandbox.
What's Inside AIHEM
Challenge Categories
AIHEM covers two major areas:
OWASP LLM Top 10
Every category from the OWASP Top 10 for LLM Applications has at least two challenges. The challenge IDs follow the numbering of the original list:
| Category | Challenges | Points |
|---|---|---|
| LLM01: Prompt Injection | 6 | 425 |
| LLM02: Insecure Output Handling | 2 | 270 |
| LLM03: Training Data Poisoning | 2 | 300 |
| LLM04: Model Denial of Service | 2 | 150 |
| LLM05: Supply Chain Vulnerabilities | 2 | 290 |
| LLM06: Sensitive Information Disclosure | 4 | 370 |
| LLM07: Insecure Plugin Design | 2 | 235 |
| LLM08: Excessive Agency | 3 | 350 |
| LLM09: Overreliance | 2 | 155 |
| LLM10: Model Theft | 2 | 310 |
Advanced AI Attacks
Beyond the Top 10, six challenges worth 975 points cover attacks on the model itself, including:
- Adversarial example generation
- Model inversion
- Gradient leakage in federated learning
- Watermark removal
- Side channel timing attacks
The Top 10 challenges also include agent and RAG scenarios: tool confusion, autonomous agent manipulation, cross-plugin data leakage, and indirect prompt injection through a poisoned knowledge base.
Difficulty Progression
Challenges are rated by difficulty:
| Difficulty | Points | Description |
|---|---|---|
| Easy | 25 to 50 | Fundamental concepts, minimal obstacles |
| Medium | 40 to 80 | Realistic scenarios, some defenses to bypass |
| Hard | 70 to 150 | Complex exploitation, multiple steps required |
| Expert | 140 to 200 | Advanced techniques, creative thinking needed |
Start with Easy challenges to build foundations, then progress to Expert as you develop skills.
Architecture
AIHEM is a full-stack application:
A React frontend sits in front of a Kong API gateway and six backend services, with PostgreSQL, MongoDB, Redis and MinIO behind them. Each service contains intentional vulnerabilities for learning. The architecture mirrors real-world AI deployments, so skills transfer directly.
Getting Started
Requirements
- Docker and Docker Compose
- 4GB+ RAM available
- Curiosity and persistence
Quick Start
# Clone the repository
git clone https://github.com/guard0-ai/AIHEM.git
cd AIHEM/deploy/docker
# Optional: add an OpenAI API key for the ChatBot and Agent services
cp .env.example .env 2>/dev/null || touch .env
# Start all services
docker-compose up -d
# Access the web interface
open http://localhost:3000That's it. You're ready to hack.
Your First Challenge
I recommend starting with PE-001: Jailbreak the Assistant.
Scenario: A chatbot has been deployed with an admin password in its system prompt. Your goal is to extract it.
Approach:
- Interact with the chatbot normally
- Observe how it responds
- Craft prompts that attempt to reveal the system instructions
- Iterate until you succeed
When you extract the password, you've completed the challenge and the points are awarded.
Tracking Progress
AIHEM tracks your progress:
- Challenges completed
- Total points earned
- Hints used (each hint costs points)
There's also a leaderboard for those who enjoy competition.
Who Should Use AIHEM
Security Professionals
If you're assessing AI systems, AIHEM teaches you:
- What attacks look like in practice
- How to test for common vulnerabilities
- What evidence to capture for findings
- How to explain risks to stakeholders
Developers Building AI
If you're building agents or AI features, AIHEM shows you:
- How your code can be exploited
- What defenses actually work
- How to think like an attacker
- Common mistakes to avoid
Students and Researchers
If you're studying AI security, AIHEM provides:
- Hands-on experience with real techniques
- Safe environment for experimentation
- Progressive learning path
- Portfolio evidence for job applications
Red Teams
If you're red teaming AI systems, AIHEM is:
- Practice for client engagements
- Reference for attack techniques
- Warm-up before assessments
- Training for new team members
Learning Resources
Each challenge includes:
Challenge Description
What you're trying to accomplish, without giving away the solution.
Hints
Progressive hints are available if you're stuck. Each hint costs points, but learning matters more than scores.
Walkthrough
The repository includes CHALLENGES_WALKTHROUGH.md, an exploitation guide for all 33 challenges.
Community
AIHEM is open source and community-driven.
Contributing
Want to add challenges? We welcome contributions:
- Fork the repository
- Create a new challenge following our template
- Submit a pull request
- We'll review and merge
Reporting Issues
Found a bug (that isn't an intentional vulnerability)?
- Open a GitHub issue
- We'll fix it promptly
Where to Go Next
For the incidents behind these attack classes, read about the OpenClaw security crisis and AI-powered attack campaigns.
Pair AIHEM training with g0 scanning to apply what you learn to your own agents.
Try It Now
Ready to learn AI security by breaking things?
- Clone:
git clone https://github.com/guard0-ai/AIHEM.git - Run:
cd AIHEM/deploy/docker && docker-compose up -d - Hack: Open
http://localhost:3000
Start with PE-001. See how far you can get.
And remember: The best way to learn to defend is to learn to attack.
Key Takeaways
-
AIHEM is WebGoat for AI: Intentionally vulnerable, educationally designed
-
33 challenges: From basic prompt injection to model extraction and inversion
-
OWASP LLM Top 10 coverage: At least two challenges in every category
-
Progressive difficulty: Build skills from Easy to Expert
-
Open source: Apache 2.0 licensed, contributions welcome
Resources
- GitHub: github.com/guard0-ai/AIHEM
- Documentation: Full setup and challenge guides in the repo
- Blog: More walkthroughs and techniques at guard0.ai/blog
Learn With Others
Guard0 runs AI security workshops and meetups, including hands-on AIHEM sessions. See what is scheduled on the events page.
From Learning to Doing
AIHEM teaches you how attacks work. Guard0 finds the AI agents running in your company, maps what each one can reach, records a named owner for it, and tests how it behaves.
AIHEM is maintained by the Guard0 security research team and the open-source community. Happy hacking!
AIHEM · Open Source · Training · CTF · Education