AI Hints
Agentic AI · Lesson 97 of 100

Quiz: Safety and Evals

The professional round.

⏱ 2 min read▶ Video at 3:01:02🎓 Free
Watch this lesson in the Agentic AI courseWatch on YouTube ↗Full Agentic AI course

Last round. These are the questions which separate someone who's built a demo from someone who can run an agent in production.

Quiz: Safety and Evals
From the video: Quiz: Safety and Evals
  • The last round
  • Demo builders
  • Production engineers

Question 1

Question 1. What is indirect prompt injection? A, A user typing an attack. B, An attack hidden in content the agent reads. C, A slow model. D, A broken tool. Pause the video and choose. The answer is B. The agent can be turned against its own user.

Quick check
What is indirect prompt injection?

Question 2

Question 2. What's the most reliable defence against it? A, A stronger prompt. B, Limiting what a tricked model can do. C, A bigger model. D, Longer instructions. Pause the video and choose. The answer is B. Assume that it's been fooled.

Quick check
What's the most reliable defence against it?

Question 3

Question 3. Which is NOT part of the lethal trifecta? A, Private data. B, Untrusted content. C, A way to send data out. D, A large model. Pause the video and choose. The answer is D. Remove any one of the three.

Quick check
Which is NOT part of the lethal trifecta?

Question 4

Question 4. Where should hard rules live? A, In the prompt. B, In code. C, In memory. D, In a summary. Pause the video and choose. The answer is B. Code can't be persuaded.

Quick check
Where should hard rules live?

Question 5

Question 5. What should memory never do? A, Store dates. B, Grant permissions. C, Have sources. D, Expire. Pause the video and choose. The answer is B. Treat it like untrusted notes.

Quick check
What should memory never do?

Question 6

Question 6. How should generated code be run? A, On your laptop. B, In a sandbox, without secrets. C, As root. D, In production. Pause the video and choose. The answer is B. No network, no keys, time limits.

Quick check
How should generated code be run?

Question 7

Question 7. What is a trajectory eval? A, Testing the tools used. B, Testing the speed. C, Testing a chart. D, Testing a model's weights. Pause the video and choose. The answer is A. Right answers can hide bad process.

Quick check
What is a trajectory eval?

Question 8

Question 8. Why repeat each eval case several times? A, To cost more. B, Outputs vary. C, It's required. D, To warm up. Pause the video and choose. The answer is B. Look at pass rates.

Quick check
Why repeat each eval case several times?

Question 9

Question 9. What's the best first stage of a launch? A, Full launch. B, Shadow mode. C, Marketing. D, A press release. Pause the video and choose. The answer is B. Nobody sees the replies.

Quick check
What's the best first stage of a launch?

Question 10

Question 10. Why put stable text first in a prompt? A, It reads better. B, Prompt caching matches from the start. C, It's shorter. D, It's safer. Pause the video and choose. The answer is B. Things that change go last.

Quick check
Why put stable text first in a prompt?

Summary

Seven or more, and you're ready for real work. Next, let's get you ready for interviews.

Key takeaways
  • Assume the model is fooled
  • Rules live in code
  • Measure everything
Next up: Interview Questions — answering like a pro