Where do you stand?

24 questions, about 18 minutes — 4 from each module. Answer honestly and skip what you don't know: we'll build a week-by-week study plan around your gaps.

You can take the diagnostic as a guest. Create a free account to save your plan and track progress against it.

Your plan settings

With an exam date we fit the plan to the weeks you have and tell you if you need more hours.

Module 1 · AI Concepts and Security Fundamentals

Question 1 Question 1 of 24
A security architect says a fraud model is 'only code', so the normal code review process is enough. What does this miss?
Question 2 Question 2 of 24
Why does overfitting create a privacy risk?
Question 3 Question 3 of 24
An investigator re-asks a chatbot the question that produced a harmful answer and gets a safe reply. What can be concluded?
Question 4 Question 4 of 24
An HR chatbot uses RAG over the HR library and answers a contractor's question with a manager's salary from a spreadsheet in the same library. Which single control would have prevented this?

Module 2 · AI Governance, Ethics, Risk and Compliance

Question 5 Question 5 of 24
Which document should state the organization's intent that all AI uses be registered before deployment?
Question 6 Question 6 of 24
An AI vendor contract allows 7 days to notify incidents. Your organization must notify CERT-In within 6 hours. What should you do?
Question 7 Question 7 of 24
A bank deploys a vendor's AI tool to screen job applicants in the EU. Under the EU AI Act, what is the bank's role and the tool's risk tier?
Question 8 Question 8 of 24
A credit model excludes gender, but approval rates still differ sharply by gender. What is the MOST likely explanation?

Module 3 · AI Data Management, Privacy and Model Security

Question 9 Question 9 of 24
When is a model registry an effective security control?
Question 10 Question 10 of 24
What distinguishes watermarking from fingerprinting of a model?
Question 11 Question 11 of 24
In differential privacy, what happens when ε (epsilon) is made smaller?
Question 12 Question 12 of 24
A model passes all accuracy tests but misclassifies any invoice containing a rare watermark as 'approved'. What attack does this suggest?

Module 4 · Secure AI Architecture, Engineering and IAM

Question 13 Question 13 of 24
An architect proposes letting the LLM decide whether a user may see a document, based on instructions in the system prompt. What is wrong?
Question 14 Question 14 of 24
Which guardrail layer decides whether a successful prompt injection achieves anything harmful?
Question 15 Question 15 of 24
An access review finds an agent identity with no owner and no activity for 90 days but broad permissions. What should happen?
Question 16 Question 16 of 24
A public inference endpoint autoscaled to 400 GPUs during a prompt-flooding attack and produced a huge bill. Which control was missing?

Module 5 · Secure AI Operations and Incident Response

Question 17 Question 17 of 24
A unique string planted in a chatbot's system prompt appears in a post on a public forum. What does this MOST likely indicate?
Question 18 Question 18 of 24
Human reviewers of an AI claims model normally override 6% of decisions. This week the rate is 0.2%. How should operations treat this?
Question 19 Question 19 of 24
Why should legal and privacy join an AI incident early?
Question 20 Question 20 of 24
Which AI weakness should usually be prioritized FIRST?

Module 6 · AI Threats, Testing and Lifecycle Management

Question 21 Question 21 of 24
An open-weight model is labeled with a permissive license but was fine-tuned from a base model with a non-commercial license. Can it be used commercially?
Question 22 Question 22 of 24
A red-team report states only 'the jailbreak succeeded'. What is missing?
Question 23 Question 23 of 24
What is the main limitation of certified robustness?
Question 24 Question 24 of 24
What distinguishes prompt injection from jailbreaking?