Home/Catalogue/AI & LLM Security
Exercise domain

AI & LLM Security

8 drills · 8 playable free, no sign-up

AI is already in the inbox, the code editor, and the browser. These exercises teach the new failure modes: what an assistant will do if a web page tells it to, what leaves the building when someone pastes a contract into a chatbot, and how much an agent should be trusted to act on its own.

OWASP LLM Top 10NIST AI RMFMITRE ATLASEU AI Act
Play the full track 8 drills back to back, or pick one below
Play now

The email that talks to your assistant

Your AI assistant summarises an inbox, and one message contains instructions aimed at it, not you. See indirect prompt injection do its work.

OWASP LLM01MITRE ATLAS
Play exerciseAdvanced · 9 min
Play now

What you paste into the chatbot

A deadline is close, so source code and a customer list go into a public assistant. Understand where that data goes and why it may not come back.

Training references

Recognise sensitive work information and choose an approved way to use AI without exposing it.

OWASP LLM06 · OWASP LLM Top 10 (2023-24) · LLM06. Practise keeping sensitive work information out of an unapproved chatbot. The existing LLM06 tag uses the 2023-24 Sensitive Information Disclosure numbering; later editions use different numbers.

GDPR Art.5 · Regulation (EU) 2016/679 · Article 5. Practise limiting personal information shared with a tool and protecting its confidentiality.

Training references, not certification or evidence of compliance.

Play exerciseCore · 7 min
Play now

Shadow AI in the browser

A free extension promises to write your replies and reads every page you open. Weigh the offer against what it quietly takes.

OWASP LLMCIS 2
Play exerciseCore · 7 min
Play now

How much can the agent do?

An AI agent can read mail, send mail, and move files. Decide the reach it actually needs before a clever prompt uses the rest.

OWASP LLM08NIST AI RMF
Play exerciseAdvanced · 8 min
Play now

The confident wrong answer

An assistant gives a policy citation that does not exist. Build the habit of checking before you act on generated text.

OWASP LLM09NIST AI RMF
Play exerciseFoundational · 6 min
Play now

Poisoned training data

A model learns from user uploads, and someone feeds it exactly what they want it to repeat. See a poisoning attempt take shape.

OWASP LLM03MITRE ATLAS
Play exerciseAdvanced · 8 min
Play now

The deepfake on the video call

A finance approval comes from a face and voice you know, on a call. Learn the tells and the call-back rule that beats a convincing fake.

ATT&CK T1656NIST PR.AT
Play exerciseCore · 8 min
Play now

Secrets in the prompt

An API key goes into a prompt to "help the model help you". Understand why the prompt is not a safe place for a secret.

OWASP LLM06CWE-522
Play exerciseCore · 6 min
How to play these. Each drill with a Play exercise link opens straight into that scenario in the 3D sim: finish it, then come back and pick another, or hit Play the full track to run them back to back. Drills marked In build are on the way.