Just Start Building Matters More Than Ever. – Hart Rossman
I'm an AI Security Engineer focused on how AI systems get attacked and how to defend them.
AI is moving faster than the security around it, and most of what’s out there is theory. I learn best by doing, so I build hands-on labs that break AI on purpose: prompt injection, poisoned tools, hijacked agents. You run them yourself, watch the attack happen, and understand how it actually works.
Everything here comes with real code you can run, fork, and adapt. Have ideas, found a better attack, or want to help? Reach out 🤝
Lab #2 tricked an AI into saying a secret. Lab #3 tricks an agent into doing something: reading your inbox, finding a password, and sending it to an attacker, all from one poisoned email. The EchoLeak pattern, on your laptop.
Lab #1 tricked a chatbot into leaking a secret. Now we give that chatbot tools through an MCP server and watch it get hijacked two new ways — a poisoned tool description, and instructions hidden in a tool's response. Build it, break it, understand it. All local, no API key.
AI is outpacing the security around it. This series is a place for security people to learn by doing. Lab #1: build a vulnerable chatbot on your laptop, trick it into leaking a secret with one sentence, and understand why the usual guards fail.
A leaked Azure storage key or Event Hub connection string is, by default, usable from anywhere on the public internet. Here's how to put that credential behind a network door — using the Azure firewall and Google's published goog.json IP ranges — with two parallel walkthroughs for the two architectures you'll actually find in the wild, plus a self-maintaining reconciler.
