Start Here
Five review stations, each a different way to practice Chapter 5. Work through them in any order. Nothing here is graded or stored, so use it as many times as you like before the quiz or the exam.
What this chapter asked you to be able to do
- Explain the role of AI in enterprise security and the risks AI systems introduce, treating AI as a defensive asset and as a new attack surface.
- Identify attacks that target AI systems, including data poisoning, model extraction, model inversion, prompt injection (direct and indirect), evasion, and jailbreaking.
- Describe data and output risks, including prompt data leakage, training data leakage, hallucinations, and bias.
- Match a control to the risk it addresses, from PII masking and rate limiting to human review, instruction isolation, and model governance.
- Evaluate the ethical and policy questions that responsible AI deployment must answer, including explainability and accountability.
Your progress this session
These counters update as you check answers on each station. They reset when you reload the page.
Key Terms Flashcards
Select a card to reveal its definition, then select again to flip it back. Try to state the definition before you flip. The counter tracks how many distinct cards you have opened.
Identifying the AI Attack
Each attack in this chapter targets a specific stage of the model lifecycle. Read each scenario and choose the attack it describes, then check your answer.
Classifying Data and Output Risks
Not every AI risk is an active attack. Read each scenario and choose which data or output risk it describes: a data privacy leak, a hallucination, or a bias and fairness problem.
Matching a Control to the Risk
Chapter 5 pairs each AI risk with the controls that reduce it. Read each goal and choose the control that best addresses it, then check your answer.
Reference: layered AI security controls
| Layer | Example control | Risk it reduces |
|---|---|---|
| Input | Mask PII before documents reach the model | Prompt data leakage |
| Input | Filter input for signs of injected instructions | Prompt injection |
| Processing | Private deployment, isolation, and rate limiting | Model extraction and inversion |
| Output | Human review of model-generated results | Hallucinations and leaked data |
| Output | Ground answers in documented sources with RAG | Hallucinations |
| Data | Vet, retain, and audit training data | Bias and data poisoning |
| Governance | Know a model's provenance and approve its data | Supply chain and accountability gaps |
One risk can be reduced at more than one layer. Prompt injection, for example, is addressed by input filtering, by instruction isolation during processing, and by output review.
Knowledge Check
Ten multiple-choice questions across the whole chapter. Answer them all, then select Score my answers for feedback and a final total.