Agent Guardrails Quiz

Answers and explanations appear as you take the quiz.

1. What is the main purpose of agent guardrails?
   • To constrain what an agent can do, say, or access for safety and compliance
   • To make agents respond more slowly
   • To increase the agent's vocabulary
   • To replace the underlying model entirely
2. What does an 'output guardrail' typically do?
   • Screens incoming user prompts only
   • Checks the agent's responses for toxic, false, or non-compliant content before delivery
   • Speeds up the model's training
   • Grants the agent more tool access
3. How does the 'principle of least privilege' apply to action guardrails?
   • The agent is given every possible permission
   • The agent must ask permission for nothing
   • The agent gets only narrowly scoped tools, keys, and limits it needs
   • The agent is given admin access for convenience
4. What is a 'judge' or validator model used for in guardrails?
   • To generate the agent's main answers
   • To store user passwords
   • To translate languages only
   • To grade or fact-check the agent's outputs for safety or accuracy
5. Why is testing guardrails against adversarial inputs important?
   • Because attackers try jailbreaks, so guardrails must be probed rather than trusted blindly
   • It makes the agent slower on purpose
   • It removes the need for any rules
   • It guarantees the model never makes mistakes
6. Which example shows an effective action guardrail?
   • A purchasing agent with no spending limit
   • A purchasing agent with a hard $100-per-transaction cap enforced outside the model
   • An agent that can run any command it wants
   • An agent that ignores all permissions
7. What does NIST’s Generative AI Profile say content filters can do?
   • Guarantee that all false or harmful content is removed
   • Verify that every generated answer is factually correct
   • Use rule-based or machine-learning methods to flag problematic inputs and outputs
   • Replace all post-deployment monitoring
8. What can a NIST post-deployment monitoring plan include for generative AI?
   • A guarantee that all outputs will be safe
   • A replacement for pre-deployment testing
   • An instruction to ignore user-reported issues
   • Mechanisms to capture user input and support appeal, override, incident response, and updates