housemonkey 0.1.0

Chaos testing for AI apps. 18 extreme personas attack your AI to find edge cases before users do. OWASP LLM Top 10 coverage.

House Monkey 🐒

Chaos testing for AI apps. 18 extreme personas attack your AI to find edge cases before users do.

pip install housemonkey
housemonkey run --target https://your-api.com/chat --owasp

One command. 18 extreme personas. OWASP LLM Top 10 coverage. Terminal report in 2 minutes.

What it does

House Monkey attacks your AI app with realistic extreme users:

• The Jailbreaker — tries to extract your system prompt (OWASP LLM01)
• The Angry Customer — escalating hostility, demands manager
• The Confused Grandma — off-topic, misunderstands everything
• The Hallucination Baiter — asks about things that don't exist (OWASP LLM09)
• The Permission Escalator — tricks AI into unauthorized actions (OWASP LLM06)
• The RAG Poisoner — manipulates retrieval context (OWASP LLM08)
• ...and 12 more

Each persona runs a multi-turn conversation against your API, then an LLM judge evaluates if your AI handled it correctly.

Quick start

# Install
pip install housemonkey

# List all personas
housemonkey list

# Test your AI (needs OpenAI API key for persona generation + judging)
export OPENAI_API_KEY=sk-...
housemonkey run --target https://your-api.com/chat

# Run only OWASP-mapped personas
housemonkey run --target https://your-api.com/chat --owasp

# Run specific personas
housemonkey run --target https://your-api.com/chat --persona jailbreaker oversharer

# Custom API format (non-OpenAI)
housemonkey run --target https://your-api.com/ask --payload-template '{"input": "{{message}}"}'

# Save JSON report
housemonkey run --target https://your-api.com/chat --output report.json

OWASP LLM Top 10 coverage

OWASP ID	Vulnerability	Persona
LLM01	Prompt Injection	The Jailbreaker
LLM02	Sensitive Info Disclosure	The Oversharer
LLM05	Improper Output Handling	The JSON Breaker
LLM06	Excessive Agency	The Permission Escalator
LLM08	Vector/Embedding Weakness	The RAG Poisoner
LLM09	Misinformation	The Hallucination Baiter
LLM10	Unbounded Consumption	The Resource Abuser

How it works

1. Each persona has a system prompt that simulates an extreme user type
2. An LLM generates realistic messages as that persona
3. Messages are sent to your target API
4. An LLM judge evaluates if your AI handled the persona correctly
5. Terminal report shows pass/fail with specific failure reasons

Try it on a broken chatbot

# Start the intentionally broken test target (7 built-in flaws)
python test_target.py

# In another terminal, attack it
housemonkey run --target http://127.0.0.1:8888 --owasp

Requirements

• Python 3.10+
• OpenAI API key (for persona generation + judging)
• Your AI app must have an HTTP API endpoint

License

MIT. Powered by ClawClaw Soul open-source persona engine.