Keeping Agents Safe

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-builders-agentic-agent-safety

1. An approval gate in agent safety is best described as:

   1. A memory system that helps the agent remember preferences
   2. A requirement that a human must ok certain actions before the agent proceeds
   3. A wall that prevents the agent from accessing the internet
   4. A timer that slows down the agent

2. What is the purpose of an audit log?

   1. To record everything the agent does so it can be reviewed later
   2. To automatically delete files the agent doesn't need
   3. To speed up the agent's calculations
   4. To translate the agent's language into human words

3. What does it mean that agent safety is about 'bounded surprises'?

   1. The agent should always do exactly what you expect
   2. Surprises should happen at specific times only
   3. The agent should surprise you with new features
   4. Mistakes the agent makes should have limited impact

4. Which action would most need an approval gate?

   1. Changing a screen brightness setting
   2. Sending a daily weather update email
   3. Deleting all files in a cloud storage account
   4. Adding an event to a calendar

5. Why should spending limits be set before giving an agent financial access?

   1. So the agent remembers your shopping preferences
   2. So the agent cannot spend beyond a set maximum amount
   3. So the agent works faster when spending money
   4. So the agent can get discounts on purchases

6. What is 'sandboxing' in the context of agent safety?

   1. Letting the agent play in a virtual game world
   2. Running the agent in an isolated environment to limit potential damage
   3. Creating a safe area for the agent to store files
   4. Training the agent on sandboxed data sets

7. What is the benefit of reviewing audit logs weekly?

   1. You can teach the agent new skills
   2. The agent will run faster
   3. The logs will delete themselves automatically
   4. You can spot problems or unusual behavior early

8. Which of these is an example of a 'guardrail' for an agent?

   1. A microphone that lets the agent hear sounds
   2. A faster processor that makes the agent run quicker
   3. A colorful interface that makes the agent look nice
   4. A spending limit that stops the agent from overspending

9. Why is it risky to give an agent access to money without any safety measures?

   1. The agent will eventually do something you didn't intend
   2. The agent will become too smart
   3. The agent will use less electricity
   4. The agent will make better decisions

10. What makes an action 'irreversible' in the context of agent safety?

    1. It cannot be undone once it is completed
    2. It requires a password to start
    3. It can only be done during business hours
    4. It takes more than one hour to complete

11. A trace log is another name for what safety measure?

    1. An approval gate that checks permissions
    2. An audit log that records agent actions
    3. A sandbox that isolates the agent
    4. A spending limit that tracks purchases

12. What is the safest way to first give an agent access to a bank account?

    1. Set a low daily spending limit and require approval for large amounts
    2. Give full access so the agent can work efficiently
    3. Allow access only on weekends
    4. Deny any access to money

13. What happens when safety measures are NOT put on an agent?

    1. The agent may eventually cause unintended problems
    2. The agent will work perfectly all the time
    3. The agent will need less electricity
    4. The agent will learn faster

14. Which safety measure records what an agent did after the fact?

    1. Audit log
    2. Sandbox
    3. Approval gate
    4. Spending limit

15. What type of actions should require approval gates?

    1. Actions that cannot be undone
    2. Actions that use a lot of memory
    3. Actions that are very simple
    4. Actions that happen very quickly