Giving an AI Agent Shell Access Without Letting It Wreck Your Machine

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-builders-agentic-ai-shell-tool-safety-r9a8-teen

1. Which safety mechanism ensures that a destructive command like rm -rf only affects a temporary environment rather than your actual computer?

   1. Sandbox isolation
   2. Command wrapping
   3. Allowlist verification
   4. Human confirmation prompt

2. An AI agent's shell tool is configured to automatically reject any command containing 'sudo'. What type of guardrail is this an example of?

   1. Resetting between sessions
   2. Confirming destructive actions
   3. Sandboxing the execution environment
   4. Allowlisting permitted commands

3. Cursor's agent mode requires you to click 'Approve' before running certain commands. What safety principle does this demonstrate?

   1. Requiring human confirmation for sensitive actions
   2. Using an allowlist to block dangerous commands
   3. Resetting the environment after each task
   4. Sandboxing the agent in a container

4. What is the main security benefit of running an AI agent in a fresh container that resets between sessions?

   1. Commands run faster without any safety checks
   2. The agent can access more system resources
   3. Any damage is limited to the temporary container and wiped clean
   4. The allowlist rules don't need to be configured

5. If you give an AI agent shell access without implementing any guardrails, what is the primary risk?

   1. The agent will run slowly
   2. The agent will refuse to work
   3. The agent will need more training data
   4. The agent could accidentally or intentionally delete important files on your actual machine

6. A developer audits their agent's shell tool and finds it has a sandbox and allowlist but no confirmation step for destructive commands. What is missing?

   1. A sandbox
   2. An allowlist
   3. Session reset capability
   4. Human confirmation for destructive actions

7. Why is it insufficient to only use an allowlist without a sandbox or confirmation step?

   1. Allowlists require expensive hardware
   2. Allowlists are too slow to be practical
   3. Allowlisted commands can still cause unintended damage if executed in the wrong environment
   4. Allowlists cannot filter file operations

8. What makes ChatGPT's code interpreter different from a directly connected shell in terms of safety?

   1. It runs in an isolated container that gets reset
   2. It cannot run any commands
   3. It asks for confirmation before every command
   4. It requires an allowlist for every command

9. A wrapped shell tool that rejects commands containing 'rm', 'mv', or 'sudo' is an example of which guardrail?

   1. Sandboxing
   2. Session resetting
   3. Allowlisting
   4. Confirmation

10. What is the purpose of giving an agent 'real power' without giving it 'the keys to your laptop'?

    1. To make the agent run faster
    2. To let the agent be useful while protecting your actual system
    3. To allow the agent to learn more quickly
    4. To reduce the cost of running the agent

11. If an agent runs inside a Docker sandbox, what happens when it executes 'rm -rf /'?

    1. The command fails because Docker blocks it
    2. The command only deletes files inside the Docker container
    3. The agent asks for confirmation first
    4. The agent deletes the host computer's entire filesystem

12. Which statement best describes why all three guardrails (sandbox, allowlist, confirmation) are needed together?

    1. They are all optional and can be used interchangeably
    2. They each require the others to function
    3. They provide overlapping protection that covers different failure scenarios
    4. Any one alone is too complicated to implement

13. Why is 'sandbox by default' recommended as a principle?

    1. Without a sandbox, any mistake has full access to your system
    2. Sandboxes are required by law
    3. Sandboxes automatically allowlist commands
    4. Sandboxes make commands run faster

14. An agent is trying to move a critical project folder to the trash. Which guardrail would most directly prevent accidental loss of important files?

    1. Human confirmation before execution
    2. Sandbox isolation
    3. Allowlist blocking the 'mv' command
    4. Session reset after the operation

15. What is the relationship between the three key terms: shell, sandbox, and allowlist?

    1. Shell is the interface, sandbox is isolation, allowlist is filtering
    2. Shell must be disabled to use sandbox
    3. Shell and sandbox are the same thing
    4. Allowlist replaces the need for a sandbox