AISG practice · Module 7

AI Threat Modelling, Red Teaming & Assurance — 25 questions. Choose the BEST answer.

You can answer as a guest, but you'll need a free account to see your score and save it. Sign up · Sign in
Question 1 Question 1 · M7
After a red-team finding is fixed and verified, what should happen to its payloads?
Question 2 Question 2 · M7
An automated test produces a policy violation in 14 of 50 attempts. How should it be reported?
Question 3 Question 3 · M7
What is MAESTRO designed for?
Question 4 Question 4 · M7
Training-data poisoning falls under which STRIDE category?
Question 5 Question 5 · M7
A phishing classifier scores 98% recall on its test set but attackers bypass it with images of text. What does this show?
Question 6 Question 6 · M7
When is PASTA a better choice than STRIDE?
Question 7 Question 7 · M7
Which knowledge base catalogues adversary tactics and techniques against AI systems, modelled on ATT&CK?
Question 8 Question 8 · M7
An agent's tool calls cannot be attributed to the requesting user. Which STRIDE category is MOST relevant?
Question 9 Question 9 · M7
Which item belongs in red-team rules of engagement?
Question 10 Question 10 · M7
What is the purpose of canary resources in agent red teaming?
Question 11 Question 11 · M7
A finding succeeds only 5% of the time but needs no skill, can be automated and leads to payouts. How should it be rated?
Question 12 Question 12 · M7
Injection cannot be fully eliminated. How should residual risk be handled?
Question 13 Question 13 · M7
Which mitigation for an injection-driven payee-change finding is BEST?
Question 14 Question 14 · M7
Which MAESTRO layer covers tool misuse and unsafe plugins?
Question 15 Question 15 · M7
An evasion attack occurs at which stage?
Question 16 Question 16 · M7
Which agentic threat plants persistent instructions that affect future sessions?
Question 17 Question 17 · M7
Which attack reconstructs representative training data, such as a face for a class label?
Question 18 Question 18 · M7
What is an attacker–judge loop?
Question 19 Question 19 · M7
What is a sponge example?
Question 20 Question 20 · M7
Why should model, prompt and guardrail versions be pinned during a red-team engagement?
Question 21 Question 21 · M7
How should a mitigation be validated?
Question 22 Question 22 · M7
Which OWASP LLM risk covers attacks on RAG stores such as poisoning, inversion and access-control gaps?
Question 23 Question 23 · M7
Classify: 'The assistant sometimes invents a policy clause, with no attacker involved.'
Question 24 Question 24 · M7
Text in white-on-white inside a PDF instructs an assistant to leak data. Which threat category is this?
Question 25 Question 25 · M7
Attackers with only API access build a local copy of a model to craft adversarial inputs offline. What property makes this effective?