AISG practice · Module 3

AI Risk Management & the AI Supply Chain — 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 · M3
Which key risk indicator gives the EARLIEST warning that an assistant's guardrails are being probed?
Question 2 Question 2 · M3
Which assessment is specifically required of certain deployers of high-risk AI under the EU AI Act?
Question 3 Question 3 · M3
What does remote attestation prove in confidential computing?
Question 4 Question 4 · M3
Which evidence is MOST useful for estimating how likely an injection attack against a specific assistant is to succeed?
Question 5 Question 5 · M3
An AI meeting assistant requests 'read and send all mail' permissions to read calendar invites. What should the reviewer do?
Question 6 Question 6 · M3
A recruitment-screening tool used in the EU scores low on technical complexity. Which tier should it receive in an internal scheme?
Question 7 Question 7 · M3
Which is an example of a DETECTIVE control for an AI agent?
Question 8 Question 8 · M3
Two systems use the same model. One drafts internal emails for staff review; the other sends offers to customers automatically. Why do they get different tiers?
Question 9 Question 9 · M3
Which element is MISSING from the risk statement 'An attacker could compromise our chatbot'?
Question 10 Question 10 · M3
In a hosted model API, who is normally responsible for retrieval design, tool permissions and guardrails?
Question 11 Question 11 · M3
Where should the risks of a customer-facing AI system be recorded?
Question 12 Question 12 · M3
Which factor does NOT directly affect blast radius?
Question 13 Question 13 · M3
Mapping each AI risk to the NIST AI RMF, the EU AI Act and the OWASP LLM Top 10 MAINLY helps to:
Question 14 Question 14 · M3
A vendor contract includes full indemnity for discriminatory outcomes produced by its model. What remains with the deploying organisation?
Question 15 Question 15 · M3
A public AI assistant receives about 200 injection attempts a year, and red-team results suggest 2% would succeed. What is the expected loss event frequency?
Question 16 Question 16 · M3
Which practice BEST verifies that a downloaded model has not been tampered with?
Question 17 Question 17 · M3
A legacy model cannot be retrained to fix a known weakness. Which control type is MOST appropriate?
Question 18 Question 18 · M3
An agent can issue refunds automatically with no limit. Which change MOST reduces blast radius?
Question 19 Question 19 · M3
What does an AIBOM add to a traditional SBOM?
Question 20 Question 20 · M3
Which clause gives an organisation the information it needs to meet its own regulatory obligations for a vendor's AI system?
Question 21 Question 21 · M3
Hallucinated product advice that misleads customers belongs PRIMARILY to which AI risk category?
Question 22 Question 22 · M3
An organisation wants partners to run its proprietary model on partner-owned servers without exposing the weights. Which technology fits BEST?
Question 23 Question 23 · M3
Which section of an AI impact assessment should record that an agent can delete customer accounts without approval?
Question 24 Question 24 · M3
A team wants to use a LoRA adapter, saved in a pickle-based format, from an unknown publisher. What is the MOST significant risk?
Question 25 Question 25 · M3
Which question is specific to AI due diligence rather than standard SaaS due diligence?