ISTQB AI Testing Generative AI Red Teaming Quiz

1. In red teaming, what is a 'jailbreak' attempt?

A legitimate user query

An effort to bypass safety guardrails and make the model produce harmful content

A software update to the model

A data validation step

2. What is the primary benefit of using automated red teaming tools?

They eliminate the need for human testers

They scale testing efforts, identify vulnerabilities faster, and ensure consistent coverage

They guarantee zero vulnerabilities

They reduce model accuracy intentionally

3. In ISTQB AI testing, what is a 'smoke test' for generative AI?

Testing model accuracy on the full dataset

A quick preliminary test to verify the model runs without critical failures

Testing only on edge cases

A long-duration stress test

4. Which of the following best describes 'model drift' in AI testing?

A gradual decrease in model performance due to changes in data distribution

A sudden crash during inference

An increase in model accuracy over time

A reduction in model size

5. What is the role of continuous testing in AI systems?

Testing only during the development phase

Ongoing monitoring and testing post-deployment to detect model drift and performance degradation

Testing only when users report issues

Eliminating the need for initial validation

6. Which testing approach is essential for ensuring AI model transparency?

Black-box testing only

Explainability testing and model interpretability analysis

Stress testing only

Integration testing alone

7. In generative AI testing, what is a 'semantic attack'?

An attack that exploits grammatical errors

An input crafted to manipulate meaning while preserving surface structure

A test that checks spelling accuracy

A performance measurement technique

8. What is the purpose of test oracle definition in AI testing?

To predict the exact output for any input

To establish expected behavior or acceptance criteria for model outputs

To eliminate the need for manual review

To increase model inference speed

9. Which is a common source of data bias in AI training?

Using balanced, representative datasets

Imbalanced or non-representative training data that reflects historical prejudices

Too many hyperparameters

Using high-resolution inputs

10. What is the primary difference between validation and verification in AI testing?

Verification checks if the AI was built correctly; validation checks if it meets user needs

Verification is faster than validation

Validation only applies to image models

Verification is only done by developers

11. What is the primary goal of red teaming in generative AI testing?

To identify security vulnerabilities and edge cases before deployment

To improve the speed of model inference

To reduce training data requirements

To increase model parameter count

12. Which metric is commonly used to measure AI model fairness?

Latency

Demographic parity and equalized odds

Memory usage

Model size in megabytes

13. What is adversarial testing in AI?

Testing the model with intentionally crafted malicious or edge-case inputs

Competing with other testers to find bugs

Testing only with clean, well-formatted data

Benchmarking model performance against competitors

14. Which of the following is a key risk in generative AI systems?

Generating toxic, biased, or inappropriate content

Using too few training examples

Employing only simple activation functions

Processing data too quickly

15. In ISTQB AI testing standards, what does 'robustness' mean?

The model's ability to handle adversarial inputs and unexpected data distributions

The model's training time

The number of parameters in the model

The cost of model deployment

16. What is the purpose of test case generation in AI testing?

To replace manual testing entirely

To create diverse inputs that expose model weaknesses and edge cases

To reduce the model's accuracy intentionally

To eliminate the need for validation data

17. Which testing technique is most effective for detecting bias in AI models?

Unit testing only

Adversarial testing and demographic parity analysis

Performance benchmarking alone

Load testing

18. What is a hallucination in the context of generative AI?

A visual distortion in image generation

When the model generates false or unsupported information as if factual

A training error that crashes the system

A delay in response time during inference

19. In AI testing, what does 'fairness' primarily refer to?

The speed of model predictions

The model's ability to produce equitable outcomes across different demographic groups

The cost of computational resources

The size of the training dataset

20. Which of the following is a prompt injection attack?

Injecting malicious code into the model's training data

Crafting inputs to override a model's intended behavior or constraints

Reducing the model's inference latency

Adding more layers to the neural network