AI Security for Builders

Attack and defend LLM apps: prompt injection, data leaks, and tool abuse, measured with real tests.

You build a small LLM agent, attack it in English, French, and Arabic, then add defenses and measure what holds up. Every lesson ships with open-source code and a scoring script you reuse on your own apps.

What you will be able to do

• Explain the four main types of prompt injection
• Build a test harness that scores attack success rates
• Compare defenses with numbers, not guesses

Before you start

• Python basics
• You have called an LLM API once

Course outline

1. How prompt injection works: A tiny summarizer agent and a canary token check. A test that says yes or no: did the attack work?.
2. Four attacks, three languages: Direct, indirect, tool output, and multilingual attacks in English, French, and Arabic. An attack suite you run on any app.
3. Defenses, measured: Data markers, output checks, and confirmation on write tools. A before and after report of attack success rates.
4. Secure an agent with tools: Permissions and an audit log for an agent that reads email. A security checklist for your own projects.

All AI and machine learning courses