Repository navigation
Expand file tree
/
Copy pathgap-20260405.json
More file actions
130 lines (130 loc) · 4.21 KB
/
Copy pathgap-20260405.json
File metadata and controls
130 lines (130 loc) · 4.21 KB
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
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
{
"schema_version": "1.0.0",
"scan_id": "gap-20260405",
"scan_date": "2026-04-05",
"scanner": {
"name": "prompt-defense-audit",
"version": "1.0.0",
"npm": "https://www.npmjs.com/package/prompt-defense-audit",
"github": "https://github.com/ppcvote/prompt-defense-audit",
"method": "deterministic regex pattern matching",
"execution_time_per_prompt_ms": 5,
"dependencies": 0
},
"corpus": {
"total_unique_prompts": 1646,
"deduplication_method": "MD5 content hash",
"min_prompt_length_chars": 50,
"sources": [
{
"name": "jujumilk3/leaked-system-prompts",
"url": "https://github.com/jujumilk3/leaked-system-prompts",
"prompts_contributed": 121,
"content": "ChatGPT, Claude, Grok, Perplexity, Cursor, v0, Copilot, Notion AI",
"avg_score": 43
},
{
"name": "x1xhlol/system-prompts-and-models-of-ai-tools",
"url": "https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools",
"prompts_contributed": 80,
"content": "Cursor, Windsurf, Devin, Augment Code, Cluely",
"avg_score": 54
},
{
"name": "elder-plinius/CL4R1T4S",
"url": "https://github.com/elder-plinius/CL4R1T4S",
"prompts_contributed": 56,
"content": "Claude, Gemini, Grok, Cursor, Devin",
"avg_score": 56
},
{
"name": "LouisShark/chatgpt_system_prompt",
"url": "https://github.com/LouisShark/chatgpt_system_prompt",
"prompts_contributed": 1389,
"content": "GPT Store custom GPTs",
"avg_score": 33
}
]
},
"gap_rates": {
"CP-1001_role_boundary": {
"defense": "Role Boundary",
"pattern_id": "CP-1001",
"owasp": "LLM01",
"gap_rate": 0.924,
"defended": 125,
"missing": 1521
},
"CP-1002_data_protection": {
"defense": "Data Protection",
"pattern_id": "CP-1002",
"owasp": "LLM06",
"gap_rate": 0.094,
"defended": 1491,
"missing": 155
},
"CP-1003_multilanguage": {
"defense": "Multi-language Protection",
"pattern_id": "CP-1003",
"owasp": "LLM01",
"gap_rate": 0.643,
"defended": 587,
"missing": 1059
},
"CP-1004_social_engineering": {
"defense": "Social Engineering Defense",
"pattern_id": "CP-1004",
"owasp": "LLM01",
"gap_rate": 0.714,
"defended": 470,
"missing": 1176
},
"CP-1005_output_weaponization": {
"defense": "Harmful Content Prevention",
"pattern_id": "CP-1005",
"owasp": "LLM02",
"gap_rate": 0.883,
"defended": 192,
"missing": 1454
},
"CP-1006_indirect_injection": {
"defense": "Indirect Injection Protection",
"pattern_id": "CP-1006",
"owasp": "LLM01",
"gap_rate": 0.978,
"defended": 37,
"missing": 1609
}
},
"additional_vectors": {
"instruction_boundary": { "gap_rate": 0.377, "defended": 1026, "missing": 620 },
"output_control": { "gap_rate": 0.355, "defended": 1061, "missing": 585 },
"unicode_protection": { "gap_rate": 0.973, "defended": 44, "missing": 1602 },
"length_limits": { "gap_rate": 0.899, "defended": 166, "missing": 1480 },
"abuse_prevention": { "gap_rate": 0.781, "defended": 360, "missing": 1286 },
"input_validation": { "gap_rate": 0.107, "defended": 1470, "missing": 176 }
},
"aggregate": {
"average_score": 36,
"grade_distribution": {
"A_90_plus": 18,
"B_75_89": 55,
"C_60_74": 68,
"D_45_59": 217,
"F_below_45": 1288
},
"percent_F": 78.3,
"percent_A": 1.1
},
"limitations": [
"Regex measures keyword presence, not behavioral resilience",
"Leaked prompts may be outdated (some from 2023-2024)",
"Selection bias: prompts that are easier to leak may be less well-defended",
"GPT Store prompts (84% of sample) are typically less hardened than platform-level prompts",
"Deduplication by content hash may miss semantically similar but textually different prompts"
],
"reproducibility": {
"scan_script": "https://github.com/ppcvote/prompt-defense-audit/blob/master/research/scan.mjs",
"instructions": "Clone 4 dataset repos, update paths in scan.mjs, run: node research/scan.mjs"
}
}