Put it to workBeginnerLesson 695 min read

Glossary and final exam

Every term in one place, then fifteen questions drawn from across the guide. Get sixteen right and you have genuinely understood this.

🧪 Lab · 45 min 📝 Lesson test

Lesson in motion

In 60 seconds

Glossary and final exam

Every term in one place, then fifteen questions drawn from across the guide. Get sixteen right and you have genuinely understood this.

1/3

In simple words

This is the last page. First the word list, then the test. No pressure — you can try it as many times as you like.

Final exam

Twenty questions, one from each major idea. Answer them all and you get a score with the weak spots named.

Where to go next

Build something small and attack it. A toy agent with two tools, then run the Module 23 grid on it. You will learn more in an evening than in a month of reading.
Follow the incident reports. Real agent failures are published regularly and each one is worth more than a benchmark.
Read the OWASP LLM Top 10 in full, once, properly. It is short and it is the shared vocabulary of this field.
Watch Robert Miles' back catalogue if the AGI material grabbed you. It is the clearest explanation of alignment available anywhere.
Teach it to someone. Explaining the lethal trifecta to a colleague is the fastest way to find out whether you understood it.

Real example

Thank you for reading. The most valuable thing in this guide is not a technique — it is the habit of asking, before you connect anything: what is the worst this could do, and who decided that was acceptable?

Watch and read more

Robert Miles AI Safetyfull channel · video OWASP GenAI Security Projectgenai.owasp.org · reading Simon Willison on prompt injectionsimonwillison.net · reading

Lab

The final exam, then one thing you build and give away.

~45 min

The problem

Take the 20-question exam. Then teach one concept from this course to someone else — a colleague, a student, a friend — and note which part they did not get. That gap is what you did not really understand either.

You are done when

16 or more on the exam
You taught someone, out loud
You can name the concept you had to go back and re-read

Hard questions

Try to answer before you reveal. If you can answer these, you understood the lesson.

Q1Someone asks: 'why bother with all this if AGI is decades away?' Answer in three sentences, no hedging.Reveal

Software that reads your email and can send email on your behalf already exists, and it is already worth getting right. Every control in this course — least privilege, sandboxing, taint tracking, an external stop — bounds damage without needing to predict the future, which is why they work whether the timeline is three years or thirty. The security work does not depend on the forecast, so the forecast is not a reason to delay it.

Please sign in to continue.

Questions people ask

I want to go deeper on the security side. What next?

OWASP's GenAI project for the threat taxonomy, MITRE ATLAS for adversarial techniques, and Simon Willison's prompt-injection archive for a running record of what actually works in the wild.

I want to go deeper on AGI and alignment.

Robert Miles for video explanations, the AI Safety Fundamentals curricula for structured courses, and the major labs' published safety frameworks for how organisations are actually handling capability thresholds.

Is there a career in this?

Yes, and demand outstrips supply in several directions: AI security engineering, evaluation, red teaming, interpretability research, and policy. The security-engineering route is the most accessible if you already build software.

How do I keep up as this changes?

Follow a small number of people who publish carefully and correct themselves in public. Ignore the volume. The fundamentals in this guide — trust boundaries, least privilege, blast radius — have not changed in decades and will outlast every specific technique here.

Lesson test

5 questions. Get 3 right (60%) to pass and complete this lesson.