Skip to content
AI Security
Practice tests · 40 items

LLM Application Security practice test

Model output as untrusted input, rendering and exfiltration channels, retrieval with real access control, vector and embedding weaknesses, knowledge-base poisoning, sensitive data in prompts and logs, guardrails and their honest evaluation, misinformation, and denial of wallet. The pool for this topic holds 11 easy, 16 medium and 13 hard items. Each answer comes with its reasoning and a link to the question or concept page it was drawn from.

Configure a test
Topic
Difficulty
Questions
11 in this pool · about 7 min
Reveal answers
Sign in to startFree account required · retakes favour new items

Sample items with answers

easy · sample
Why should an application treat model output as untrusted input?
Models often misspell
✓Injected text can shape it, so it needs the same checks as user input
Model output is signed by the provider, so altering it would break the signature
Output from a model is encrypted and must be decrypted before it can be used safely

Whatever reaches the context, including attacker-controlled documents, can influence what the model writes. If that output goes into HTML, SQL, a shell or a file path without validation and encoding, the attacker gets a classic injection by way of the model. OWASP lists this as LLM05.

easy · sample
How can a Markdown image in a chatbot answer leak data without a single click?
✓The client fetches the image URL, and data in its query string goes to that server
Images contain hidden scripts
The image is shown to other users of the chatbot, who can read the data in it
Markdown images upload the user's local files to the image host when they are rendered

If injected instructions make the model write an image whose URL contains conversation data, the browser requests it automatically and the attacker's server logs the data. Several assistants were vulnerable to this. Blocking images from non-allowlisted hosts closes the channel.

After the test

Choosing from options is easier than reasoning it out in front of a panel. These pages make you reason it out: