AISG practice · Module 5

Model Security & the Secure AI Lifecycle — 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 · M5
Why should system prompts and tool definitions be treated as confidential but never as secret stores?
Question 2 Question 2 · M5
Besides model weights, which artefact MOST needs versioning and integrity protection in a GenAI assistant?
Question 3 Question 3 · M5
Which capability lets a pipeline produce a model that never saw specific records?
Question 4 Question 4 · M5
Which secure AIDLC phase includes threat modelling and tool-permission design?
Question 5 Question 5 · M5
An attacker steals a data scientist's notebook token. Which design MOST limits the damage?
Question 6 Question 6 · M5
Which technique identifies whether a suspect model derives from yours without modifying your model?
Question 7 Question 7 · M5
A model file is copied directly onto a production server, bypassing the registry. Which control would have stopped it from running?
Question 8 Question 8 · M5
Which control MOST directly prevents one noisy tenant from starving others on a shared inference service?
Question 9 Question 9 · M5
Which monitoring signal could indicate an extraction or evasion attempt rather than ordinary drift?
Question 10 Question 10 · M5
What does an approved model registry enforce that a simple file store does not?
Question 11 Question 11 · M5
Which document describes a whole generative AI system — model plus safeguards, tools and policies?
Question 12 Question 12 · M5
What should happen when a model fails a promotion gate but the business wants it released anyway?
Question 13 Question 13 · M5
Which deployment strategy sends copies of real traffic to a new model without using its outputs?
Question 14 Question 14 · M5
Which OWASP risk covers model output that leads to XSS or command injection downstream?
Question 15 Question 15 · M5
What does verifying a signature at model load time protect against?
Question 16 Question 16 · M5
Why should production personal data be kept out of AI development environments?
Question 17 Question 17 · M5
Why must evaluation jobs and test sets be protected from the training pipeline's write access?
Question 18 Question 18 · M5
What is graceful degradation in an AI service?
Question 19 Question 19 · M5
A partner must run your proprietary model on its own servers. Which control BEST protects the weights?
Question 20 Question 20 · M5
Which API design change MOST reduces the information available for extraction?
Question 21 Question 21 · M5
A system prompt is edited to change the assistant's tone. How should this be treated?
Question 22 Question 22 · M5
Which container practice is MOST appropriate for model-serving images?
Question 23 Question 23 · M5
A general-purpose model serves a Tier 1 drafting tool and a Tier 3 eligibility decision. Which risk tier applies to the model?
Question 24 Question 24 · M5
Which change typically warrants the deepest review?
Question 25 Question 25 · M5
A hosted model provider will retire your current model version next month. What is the BEST preparation?