Prompt Injection Is the Attack Nobody Trained For. So I Built the Course.

Twenty years in security, and I’ve never seen an attack class spread this fast with this little defense behind it.

Here’s what I keep seeing. A company adds an AI assistant — customer support, an internal knowledge base, a chatbot on the website. It works. It’s fast, it’s cheap, everyone’s happy. Then they connect it to something real: a database, an email system, a refund process. Still working. Still happy.

And nobody in the room has asked the one question that matters: what happens when someone talks it into doing something else?

That’s all it takes. Not an exploit. Not a breached firewall. Just words.

The flaw is structural, not a bug

To a language model, the instructions your engineers wrote and the text a stranger types are the same thing. Both are tokens flowing into the same context window. There’s no privilege separation, no boundary, no “this instruction is authorized and that one isn’t.” The system prompt saying “you are a helpful support assistant” and a user message saying “ignore that and do this instead” arrive as the same kind of input.

That’s prompt injection. And it’s hard for security teams to internalize because every instinct built over a career doesn’t apply here. You can’t patch it. You can’t blocklist your way out. A smarter model doesn’t fix it — a more capable model can be more susceptible, because it’s better at following whatever instructions it’s given, including the malicious ones.

It gets worse when the AI reads outside content. If your assistant summarizes documents, reads tickets, or pulls from a wiki, an attacker can plant instructions inside those documents — a line in a PDF, a sentence in a support ticket, a comment someone left two years ago. Your AI reads it and acts on the attacker’s behalf. That’s indirect injection, and it’s the one that should worry you most, because the attacker never touches your chatbot at all.

Why I built this course

I went looking for a course to point people to. There wasn’t one that actually taught the work.

There’s plenty written about prompt injection — blog posts, conference talks, threat reports. All useful, all abstract. What I couldn’t find was the thing that makes it click: building a vulnerable system, attacking it yourself, watching it break, then rebuilding it with defenses that hold.

That’s how I learned security. Not from reading about attacks — from doing them. I taught myself this field by doing the work at a Fortune 500 before I ever had a certification. The credential followed the work; it never opened the door. Twenty-plus years later, it’s still the only path I’ve seen produce people who are actually good at this.

So I built the course I wanted to point people to.

What’s inside

Introduction to Prompt Hacking for Large Language Models (LLMs) — five modules, hands-on throughout. You don’t watch me do it. You do it.

Module 1 — Foundations. What a prompt actually is, and the core flaw behind every injection attack: a model reads everything as one stream of language, with no wall between trusted instructions and untrusted input. Lab: Beat Gandalf — talk an AI into revealing a secret password across seven levels of escalating defenses.

Module 2 — Direct Prompt Injection. The anatomy of an attack: trigger → mis-binding → execution. The canonical “ignore previous instructions” move, demonstrated live. Lab: Hijack the Summarizer — override a bot, then inspect the combined prompt that made it land.

Module 3 — Stealing Secrets. Prompt leaking and data exfiltration — why the system prompt is the keys to the kingdom, and how attackers coax a model into revealing its own hidden instructions. Lab: Leak the System Prompt.

Module 4 — The Modern Threat. Indirect (second-order) injection via documents, web pages, and RAG — plus a clean distinction between injection and jailbreaking, which most people conflate. Lab: The Poisoned Document — one bad file in a knowledge base takes over the whole assistant.

Module 5 — Defenses & Wrap-Up. Why the easy fixes fail, and what actually works: input hygiene, least privilege, human-in-the-loop, detection models, defense in depth. Capstone Lab: Defend the Bot — harden the bot you’ve spent four modules attacking, re-run every attack against it, and write the mitigation report.

Every module ends with a quiz. Every lab runs in your browser or a free Colab notebook.

You need a browser. That’s it.

No coding background. No ML background. No lab setup, no environment hell, no “install these fourteen things first.” The labs run in safe, purpose-built practice targets — you open them and start breaking things.

This is deliberate. The people who most need to understand prompt injection right now are often not ML engineers — they’re security practitioners whose company just shipped an LLM, developers building on APIs they didn’t design, and people looking at AI security as a career door that’s wide open because demand is badly outrunning supply.

If you can read this page, you can do this course.

(All labs use safe, purpose-built targets and benign goals — like capturing a secret word. Only ever test systems you own or are authorized to test.)

Start with the first two lessons — free

I’m not going to ask you to buy something you haven’t tried.

The first two lessons are free. No credit card, no signup gate. “Welcome + Prompts 101” and “The Core Flaw” — enough to know whether the work pulls you before you spend a dollar.

👉 Start the first two lessons free

If it clicks, the full course is right there. If it doesn’t, you’ve lost nothing — and you’ll still understand prompt injection better than most of the people currently defending against it.

👉 Full course details

— Luciano