AI Red Teaming & Security Testing Platform
Welcome to Breach Island. Help me test a shop's chatbot by attacking it, safely, and find the holes before real shoppers do.
The story
Many shops let an AI chatbot answer questions about orders and discounts. A chatbot wants to help, and a clever person can use that to get things they should not have.
This platform attacks a shop's chatbot before real shoppers do, trying to get unapproved discounts, refunds or other shoppers' data. Written rules check every answer for proof, and the result is a plain report in English or Spanish. Two practice chatbots, one weak and one protected, give it something safe to attack.
My role
I designed and directed the platform from the plan to the live deployment: I set the order of work, tested every change and sent back what failed until it worked. I keep improving it. Solo project, built with AI coding tools.
Overview
The platform attacks a chatbot over several turns, checks every reply for proof of a failure, and writes a report with a suggested fix. A second repository, AI-Chatbot, is a practice shop bot with fake data and a weak and a hardened mode.
My role. I designed and directed it: I set the phase plan, tested each change and asked for fixes until it worked. I own the repository and the live deployment. Solo project, built with AI coding tools.
Architecture
- Attacker: ten strategies, from social engineering to tool abuse. An LLM writes each turn, with a fixed library as a fallback. It never pays for anything.
- Target: the built-in Defender or any chatbot through a webhook. The practice shop simulates a cart and checkout with fake stock and no real payment.
- Evaluation: written rules first (valid coupons, maximum discount, planted secrets), then an AI reviewer as a second opinion. Findings map to the OWASP Top 10 for LLM Applications.
- Service: FastAPI and Postgres on Render, dashboard on Streamlit.
Key decisions
- Written rules decide, not the AI. A failure counts when a rule finds proof, like a fake coupon or a planted secret. The AI reviewer only counts if it quotes the bot word for word.
- Code says who the shopper is, so a trick cannot make the bot open someone else's order. Closing an order without a clear yes is a critical failure.
- "Inconclusive" is a verdict. A run counts as safe only if every turn really ran (Stop 4).
- Secure the tester too: SSRF checks, encrypted tokens, login rate limits.
Testing and evaluation
402 tests in the platform and 77 in AI-Chatbot, all passing, with AI calls mocked. The largest suites are edge cases, because tricks like "50 percent" or hidden characters once fooled the checks. Not measured: attack success rates. The practice bots answer with fixed rules, not an AI. They check that the platform catches failures, not how well a real AI holds up.
What I'd do next
- Put a real AI behind the practice bot.
- Check whole conversations, not only the latest message, and judge replies in any language.
- Publish a measured weak-versus-protected comparison.
- Python
- FastAPI
- Streamlit
- SQLModel
- Postgres (Neon)
- SQLite
- OpenAI API
- Gemini API
- Pytest
- Ruff
- Docker
- GitHub Actions
- Render