CAIC Cybersecurity & Risk 2 — Questions and Answers
Question 1: Which attack vector specifically targets the training data of an AI model to degrade its performance or introduce malicious behavior?
- Prompt injection
- Data poisoning (Correct answer)
- Model inversion
- Adversarial evasion
Correct answer: Data poisoning
Data poisoning corrupts training data to manipulate model outputs or degrade accuracy.
Question 2: An AI system used for loan approvals is found to systematically deny applications from a protected demographic. Which risk category best describes this?
- Cybersecurity risk
- Algorithmic bias risk (Correct answer)
- Operational availability risk
- Regulatory compliance risk
Correct answer: Algorithmic bias risk
Algorithmic bias risk covers discriminatory outcomes embedded in model outputs affecting protected groups.
Question 3: What is the primary purpose of an AI model's threat model during the risk assessment phase?
- To train the model on threat intelligence feeds
- To identify assets, adversaries, and attack surfaces relevant to the AI system (Correct answer)
- To configure firewall rules for the inference server
- To audit user access logs post-deployment
Correct answer: To identify assets, adversaries, and attack surfaces relevant to the AI system
A threat model maps the assets, potential adversaries, and attack surfaces to guide security controls.
Question 4: A prompt injection attack against an LLM-powered customer service bot successfully makes it reveal internal system instructions. Which control would most directly mitigate this?
- Encrypting the model weights
- Implementing output filtering and strict prompt sandboxing (Correct answer)
- Using a smaller model with fewer parameters
- Enabling multi-factor authentication for end users
Correct answer: Implementing output filtering and strict prompt sandboxing
Output filtering and prompt sandboxing limit the model's ability to expose sensitive instructions or be manipulated via crafted inputs.
Question 5: Under NIST AI RMF, which function focuses on detecting adverse events or anomalies in an AI system's behavior?
- Govern
- Map
- Measure (Correct answer)
- Manage
Correct answer: Measure
The Measure function in NIST AI RMF focuses on analyzing and monitoring AI risks, including detecting anomalies.
Question 6: Which technique allows a red team to assess an AI system's robustness by systematically crafting inputs designed to cause misclassification?
- Transfer learning evaluation
- Adversarial testing (Correct answer)
- Hyperparameter tuning
- Feature importance analysis
Correct answer: Adversarial testing
Adversarial testing involves crafting perturbed inputs to probe model vulnerabilities and misclassification thresholds.
Question 7: An organization deploys an AI chatbot that retains conversation history across sessions without user consent. Which risk is most prominently violated?
- Model drift risk
- Data privacy and retention risk (Correct answer)
- Supply chain risk
- Availability risk
Correct answer: Data privacy and retention risk
Storing user conversation data without consent violates data privacy regulations such as GDPR and CCPA.
CAIC Cybersecurity & Risk 2
Which attack vector specifically targets the training data of an AI model to degrade its performance or introduce malicious behavior?