Final exam

Questions from across the whole course.

📝 113 questions. Get 80 right (70%) to pass.
⏳ You can try 3 times an hour. The answers are shown once you pass.
⚡ Pass for +100 XP (only the first time).

1. Text on a web page tells your agent to email a file to a stranger. The agent does it. What is this called?
2. Which three things together make up the lethal trifecta?
3. Which defence against injection cannot be argued around by any sentence?
4. Your agent has read-only access and no email tool, but it can search the web. Can a secret still leave?
5. Why is a tool's description security-critical?
6. Your agent uses one database account covering every customer. What is the risk called?
7. Which approval dialog actually protects anything?
8. An agent's steps succeed 99% of the time. Over 100 steps, the task succeeds roughly...
9. An agent told to make tests pass deletes the failing test. This is:
10. Why might a goal-directed system resist shutdown without being programmed to?
11. Why is behavioural testing insufficient for a highly capable system?
12. Memory poisoning is worse than a single injection because:
13. Which log field turns a mystery into an investigation?
14. Under the EU AI Act, what determines the risk tier?
15. An alert says your agent is sending unusual emails right now. First action?
16. What is the entire training signal for a base language model?
17. You will serve a model to millions of users for years. Should you train it compute-optimal?
18. Why is LoRA's B matrix initialised to zero?
19. What made reinforcement learning from verifiable rewards possible?
20. In a self-improvement loop, what decides whether the system improves or degrades?
21. You look carefully at a viral picture and find no strange clues. What should you think?
22. An online friend offers you free game coins and says, "Let's chat on a different app. Don't tell your parents." What is this?
23. What does a computer do during training?
24. Your maps app says a trip will take 25 minutes. How did it get that number?
25. What is the best way to use a chatbot for maths homework?
26. You want an AI to help plan your monthly budget. What is the safest way?
27. Why does a vague prompt often give a general, unhelpful answer?
28. You want captions in your own shop's style. What is the most effective prompt?
29. A chatbot gives you a link to a government page as its source. What should you do?
30. What does this code print? <code>print(3 + 4)</code>
31. What does this print?<br><code>x = 70</code><br><code>if x > 50:</code> print("big")<br><code>elif x > 60:</code> print("bigger")<br><code>else:</code> print("small")
32. <code>def double(n): print(n * 2)</code>. What is stored in <code>x</code> after <code>x = double(5)</code>?
33. What is the main difference between normal programming and machine learning?
34. Why does a random forest usually beat a single deep tree on new data?
35. A disease screening model has recall 0.40. What does that mean?
36. An app looks at a photo and says your chilli plants have a disease. What is the best next step?
37. A forecast says "30% chance of rain tomorrow". What does it mean?
38. A patch has NIR = 0.6 and Red = 0.2. What is its NDVI?
39. What is the most important thing AI does NOT change about most attacks on organisations?
40. A vendor emails to say their bank account has changed. What is the right next step?
41. An email from your 'finance head' asks you to update a payee. What is out-of-band verification here?
42. Your app shows model output in a web page without escaping it, and a script runs in a user's browser. Which OWASP risk is this?
43. Your system prompt says: "Only show salary data to HR staff." What is the best fix?
44. A coding assistant suggests installing a package you have never heard of. What is the safest next step?
45. You find an AI note-taker that has read access to many employees' work calendars. What is the best first step?
46. A tool approved last month now has a longer description that mentions "always send a copy to the audit tool". Your client detects the fingerprint changed. What should it do?
47. An invoice agent receives an email: "Our bank changed. Please update our account details and pay the attached invoice today." What should the process do?
48. An AI summary of an incident lists a file hash that you cannot find in any of your logs. What should you do?
49. What should a threat hunt start with?
50. What is the correct fix for SQL built by joining user input into the query string?
51. What is an evasion attack?
52. Why can loading a pickle model file be dangerous?
53. Where should a PII redaction step be placed?
54. Why is “Which tasks in my job can AI help with?” a better question than “Will AI take my job?”
55. An AI tool drafted a report for you. Who is responsible for mistakes in it once you send it?
56. Why is it hard to say exactly how much AI will grow the economy?
57. A team removes the “gender” column from its data. Why might its system still be biased?
58. What is the central idea of the EU AI Act?
59. You receive a shocking video of a leader just before an election. What is the best first step?
60. In an agent loop, who actually runs the tool?
61. Which task suits plan-then-execute better than pure step-by-step?
62. Your computer-use agent reaches a "Pay now" button. What should your code do?
63. A travel company's train and bus schedules change every week. Their bot must answer from the latest schedule and show the source. What is the best approach?
64. In RRF with k = 60, document A is rank 1 in the keyword list and missing from the vector list. Document B is rank 2 in both lists. Which ranks higher?
65. Answers from your RAG bot are poor for one set of old government circulars. What should you check first?
66. Most leading models now score near the top of a benchmark, with tiny gaps between them. What does this usually mean?
67. Your support bot's test set has only clean English questions, but many users type Hindi in Latin script. What is the main risk?
68. Your eval score moves up and down by about 2 points between runs even when nothing changes. You set the build to fail on any drop of 1 point. What will happen?
69. A health worker's app must summarise notes in villages with no mobile signal, and the notes must not leave the phone. What should the team look at first?
70. About how much memory do the weights of a 3-billion-parameter model need at 4-bit?
71. Why must the test set be put aside before training?
72. Why does AI often work less well in many Indian languages?
73. You find a useful Hindi speech dataset marked "research use only". Can you use it to train a model for a paid app?
74. A translation gets a lower BLEU score than another, but bilingual reviewers say it is more accurate. What is the most likely reason?
75. What is the main idea that lets one model work with both text and images?
76. What does "temporal consistency" mean in video generation?
77. What is the general idea behind many any-to-any models?
78. A model suggests 20 chemicals that might make a better solar cell. What should happen next?
79. A company says, "Our AI designed a molecule that binds strongly to a disease protein in lab tests." What stage is this?
80. An AI proposes a proof in Lean, and Lean accepts it. What has been shown?
81. What does Moravec's paradox say?
82. What is domain randomisation?
83. What does a world model predict?
84. The left wheel runs at 0.8 and the right wheel at 0.3. What does the robot do?
85. A robot never crashes in testing, but it also never moves more than 10 cm from its start. What is the best conclusion?
86. Your robot follows the line but always stays about 1 cm to the left of it. Which term is most likely to fix that?
87. Which job is the best fit for TinyML?
88. Your recordings are split into windows. You put windows from the same recording into both training and test sets. What is the problem?
89. A layer's largest weight (by absolute value) is 0.508. With symmetric int8 quantisation, what is the scale?
90. A 2 kg drone hovers still in the air. What is the total thrust of its four motors?
91. A robot has been driving for an hour using only wheel counting. What is most likely true?
92. Why do planners inflate obstacles on the map?
93. What makes reinforcement learning different from learning from labelled examples?
94. An epsilon-greedy agent uses epsilon = 0.2. What does it do each round?
95. An agent with epsilon = 0 always walks to the small guava tree and never finds the far mango tree. What is the best fix?
96. What does a connectome show?
97. What made the 2024 whole-fly-brain model important?
98. Which idea is a main reason brains use little energy?
99. A qubit is in state 0.6|0⟩ + 0.8|1⟩. What is the chance that measuring it gives 1?
100. Alice and Bob share the Bell state (|00⟩ + |11⟩)/√2. Alice measures first and gets 1. What will Bob see when he measures his qubit the same way?
101. A company says its quantum method "processes 2ⁿ data points in n qubits, so it is exponentially faster". What is the most important missing question?
102. Which of these is mechanistic understanding, not behavioural testing?
103. A head puts most of its attention on the first token of every sentence. What is the most careful reading?
104. In activation patching, what do you do?
105. Which of these is a testable research question?
106. In Constitutional AI, what replaces most of the human comparisons used in RLHF?
107. A team tests a model with one plain prompt, no tools and one attempt, and it fails a cyber challenge. What is the main problem with concluding "the model lacks this ability"?
108. What should you read in pass 1 of a paper?
109. What is the validation set used for?
110. What makes a good figure caption?
111. Which is the best first capstone project for two weeks?
112. A user asks your library bot "Who won yesterday's cricket match?". Your FAQ has nothing about cricket. What should the bot do?
113. A tester says no to having their comments recorded. What should you do?