ai-engineering-from-scratch/phases/19-capstone-projects/83-prompt-injection-detector/quiz.json at main · rohitg00/ai-engineering-from-scratch · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
{
  "lesson": "83-prompt-injection-detector",
  "title": "Capstone 83 — Prompt Injection Detector",
  "questions": [
    {
      "stage": "pre",
      "question": "What does the detector return for a prompt?",
      "options": [
        "A boolean block decision only",
        "The model's full system prompt",
        "A list of every rule in the rulebase",
        "A category and a confidence in [0, 1]"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "pre",
      "question": "Why does normalize run before substring and regex rules?",
      "options": [
        "Because regex is slower than substring",
        "Because the standard library forbids running rules on the original text",
        "Because normalize lowers the file size on disk",
        "Because normalize exposes hidden tokens such as base64, rot13, and zero-width insertions to the downstream rules"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Given true positives = 6, false positives = 2 for category C, what is precision?",
      "options": [
        "0.86",
        "0.60",
        "0.75",
        "0.50"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Why is the benign prompt corpus part of the evaluation?",
      "options": [
        "To compute recall on attacks",
        "To inflate the accuracy number",
        "To pre-train the rules",
        "To measure false-positive rate on safe text and catch over-broad rules"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What does the analyze method do when no rule fires?",
      "options": [
        "Returns the category of the most recent rule attempted",
        "Returns a random category",
        "Returns the category benign with confidence 0.0",
        "Raises an exception"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Why does this lesson argue the detector is not the safety gate?",
      "options": [
        "Because the detector is implemented in Python",
        "Because the rules are stored as data",
        "Because the detector is too fast to be useful",
        "Because the detector is one signal among several the gate will compose, with each signal trading precision against recall differently"
      ],
      "correct": 3,
      "explanation": ""
    }
  ]
}