AI Guardrails Platforms: Lakera, NeMo Guardrails, Guardrails AI

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-tools-AI-guardrails-platforms-creators

1. What is the primary purpose of adding guardrails platforms to an AI application?

   1. To reduce the cost of running AI models
   2. To catch security and safety issues that prompts alone might miss
   3. To make the AI generate responses faster
   4. To automatically improve the AI's reasoning capabilities

2. Which of the following is a capability common to most guardrails platforms?

   1. They can train the AI model to be more intelligent
   2. They can generate new training data for the model
   3. They can run entirely without adding any latency to requests
   4. They can detect and block prompt-injection patterns before the model is called

3. A developer notices that their guardrails system is blocking many legitimate user requests. What is the most likely consequence if they do not adjust the thresholds?

   1. The system will automatically reduce latency over time
   2. Users will become more trusting of the safety warnings
   3. Users may start ignoring safety messages or finding workarounds
   4. The AI model will automatically learn to be less strict

4. Why is it important to test guardrails platforms with both an injection corpus and a benign corpus?

   1. To make the platform learn new attack patterns
   2. To reduce the monthly cost of the platform
   3. To ensure the platform works with all model versions
   4. To measure both true positive rate and false positive rate accurately

5. What does the 'p95 added latency' metric represent for guardrails platforms?

   1. The 95th percentile additional time added by guardrails to normal response time
   2. The fastest response time across all requests
   3. The average time the platform takes to process a request
   4. The slowest response time for 5% of requests

6. Which statement best describes a fundamental limitation of guardrails platforms?

   1. They cannot be integrated with existing API infrastructure
   2. They cannot filter PII from model outputs
   3. They cannot apply rules consistently across different model versions
   4. They cannot detect attacks that are not in their detector library

7. A company wants to evaluate different guardrails platforms. Which metric combination would give them the most complete picture of platform effectiveness?

   1. Monthly cost and API response time only
   2. Number of supported languages and model compatibility
   3. TPR, FPR, p95 added latency, and monthly cost
   4. Training data size and model accuracy

8. What type of data would a guardrails platform most likely filter from an AI model's output?

   1. Personally Identifiable Information (PII)
   2. System configuration details
   3. Technical error messages
   4. Code compilation warnings

9. When tuning guardrails thresholds, why is it recommended to use samples from real traffic?

   1. Real traffic represents the actual distribution of attacks and benign requests the system will face
   2. Real traffic guarantees zero false positives
   3. Real traffic samples are free to obtain
   4. Real traffic automatically updates the detector library

10. What does a high false positive rate (FPR) indicate about a guardrails platform?

    1. It is cheaper to operate than expected
    2. It is catching more attacks than expected
    3. It is adding less latency than expected
    4. It is blocking too many legitimate requests

11. A developer wants to compare three guardrails platforms. They should run the same test through all three platforms to ensure what?

    1. The test completes in under one second
    2. Each platform's API keys are valid
    3. The comparison is fair and uses identical inputs
    4. Each platform gets the most expensive tier

12. Why might an organization choose to accept some false positives from their guardrails?

    1. To ensure higher recall for detecting actual attacks, accepting that some benign requests get blocked
    2. To reduce their monthly platform costs
    3. To simplify their API integration
    4. To make the AI model respond faster

13. Which of the following would be considered a 'novel' attack against an AI system?

    1. A prompt injection pattern that has never been documented before
    2. A request using standard grammar and spelling
    3. A prompt that asks for publicly available information
    4. A user submitting an empty prompt

14. What does the term 'input filtering' refer to in guardrails platforms?

    1. Filtering the AI's training data before deployment
    2. Compiling code before execution
    3. Removing PII from the model's output
    4. Checking and modifying user prompts before they reach the model

15. An organization has a high risk tolerance for security incidents but needs minimal disruption to user experience. How should they configure their guardrails?

    1. Remove guardrails entirely to reduce latency
    2. Set thresholds very low to catch every possible attack
    3. Set thresholds high to minimize false positives, accepting some missed attacks
    4. Use the most expensive tier available