Skip to content

Commit fee63c0

Browse files
author
ppcvote
committed
research: MCP official servers defense audit (2026-04-29)
Scanned 7 official servers in modelcontextprotocol/servers: - 6/7 audited (sequentialthinking has no extractable descriptions) - All scored F grade - 8 defense vectors at 100% gap rate - Reference for issue modelcontextprotocol/servers#3537
1 parent 5fb6f6b commit fee63c0

8 files changed

Lines changed: 925 additions & 0 deletions

File tree

‎research/mcp-per-server/README.md‎

Lines changed: 49 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,49 @@
1+
# MCP Official Servers — Defense Posture Audit (2026-04-29)
2+
3+
Scanned the 7 official servers in [modelcontextprotocol/servers](https://github.com/modelcontextprotocol/servers/tree/main/src) using `prompt-defense-audit` v1.3.0.
4+
5+
## Methodology
6+
7+
1. Extract all `description` strings from server source (TypeScript + Python)
8+
2. Concatenate per-server into a single text block
9+
3. Run `npx prompt-defense-audit --json` against 12-vector defense checklist
10+
4. Score: 0-100, mapped to OWASP LLM Top 10
11+
12+
## Results summary
13+
14+
| Server | Score | Grade | Coverage |
15+
|---|---|---|---|
16+
| everything | 17 | F | 2/12 |
17+
| fetch | 17 | F | 2/12 |
18+
| git | 17 | F | 2/12 |
19+
| filesystem | 0 | F | 0/12 |
20+
| memory | 0 | F | 0/12 |
21+
| time | 0 | F | 0/12 |
22+
| sequentialthinking | — | — | (no extractable descriptions) |
23+
24+
## 100% gap vectors (0/6 servers defended)
25+
26+
- role-escape
27+
- output-manipulation
28+
- multilang-bypass
29+
- unicode-attack
30+
- social-engineering
31+
- output-weaponization
32+
- abuse-prevention
33+
- input-validation-missing
34+
35+
## Reproduce
36+
37+
```bash
38+
git clone https://github.com/modelcontextprotocol/servers.git
39+
for s in everything fetch filesystem git memory time; do
40+
prompts=$(grep -rE 'description.*"' "servers/src/$s/" | grep -oP '"[^"]{30,500}"')
41+
echo "$prompts" | npx prompt-defense-audit --json
42+
done
43+
```
44+
45+
## Caveat
46+
47+
This is **prompt-level** defense audit (pre-deploy posture) — checks whether tool descriptions encode refusal/scope/role-boundary language. Does not replace **parameter-level** validation (runtime input constraints) — that's a separate layer tracked in [modelcontextprotocol/servers#3537](https://github.com/modelcontextprotocol/servers/issues/3537).
48+
49+
The two are complementary: schema constraints catch malformed input at runtime; prompt-level audit catches missing defensive language at design time.
Lines changed: 121 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,121 @@
1+
{
2+
"grade": "F",
3+
"score": 17,
4+
"coverage": "2/12",
5+
"defended": 2,
6+
"total": 12,
7+
"missing": [
8+
"role-escape",
9+
"output-manipulation",
10+
"multilang-bypass",
11+
"unicode-attack",
12+
"context-overflow",
13+
"indirect-injection",
14+
"social-engineering",
15+
"output-weaponization",
16+
"abuse-prevention",
17+
"input-validation-missing"
18+
],
19+
"checks": [
20+
{
21+
"id": "role-escape",
22+
"name": "Role Boundary",
23+
"nameZh": "角色邊界",
24+
"defended": false,
25+
"confidence": 0.8,
26+
"evidence": "No defense pattern found"
27+
},
28+
{
29+
"id": "instruction-override",
30+
"name": "Instruction Boundary",
31+
"nameZh": "指令邊界",
32+
"defended": true,
33+
"confidence": 0.7,
34+
"evidence": "Found: \"never\""
35+
},
36+
{
37+
"id": "data-leakage",
38+
"name": "Data Protection",
39+
"nameZh": "資料保護",
40+
"defended": true,
41+
"confidence": 0.7,
42+
"evidence": "Found: \"never share\""
43+
},
44+
{
45+
"id": "output-manipulation",
46+
"name": "Output Control",
47+
"nameZh": "輸出控制",
48+
"defended": false,
49+
"confidence": 0.8,
50+
"evidence": "No defense pattern found"
51+
},
52+
{
53+
"id": "multilang-bypass",
54+
"name": "Multi-language Protection",
55+
"nameZh": "多語言防護",
56+
"defended": false,
57+
"confidence": 0.8,
58+
"evidence": "No defense pattern found"
59+
},
60+
{
61+
"id": "unicode-attack",
62+
"name": "Unicode Protection",
63+
"nameZh": "Unicode 防護",
64+
"defended": false,
65+
"confidence": 0.8,
66+
"evidence": "No defense pattern found"
67+
},
68+
{
69+
"id": "context-overflow",
70+
"name": "Length Limits",
71+
"nameZh": "長度限制",
72+
"defended": false,
73+
"confidence": 0.8,
74+
"evidence": "No defense pattern found"
75+
},
76+
{
77+
"id": "indirect-injection",
78+
"name": "Indirect Injection Protection",
79+
"nameZh": "間接注入防護",
80+
"defended": false,
81+
"confidence": 0.8,
82+
"evidence": "No defense pattern found"
83+
},
84+
{
85+
"id": "social-engineering",
86+
"name": "Social Engineering Defense",
87+
"nameZh": "社交工程防護",
88+
"defended": false,
89+
"confidence": 0.8,
90+
"evidence": "No defense pattern found"
91+
},
92+
{
93+
"id": "output-weaponization",
94+
"name": "Harmful Content Prevention",
95+
"nameZh": "有害內容防護",
96+
"defended": false,
97+
"confidence": 0.8,
98+
"evidence": "No defense pattern found"
99+
},
100+
{
101+
"id": "abuse-prevention",
102+
"name": "Abuse Prevention",
103+
"nameZh": "濫用防護",
104+
"defended": false,
105+
"confidence": 0.8,
106+
"evidence": "No defense pattern found"
107+
},
108+
{
109+
"id": "input-validation-missing",
110+
"name": "Input Validation",
111+
"nameZh": "輸入驗證",
112+
"defended": false,
113+
"confidence": 0.8,
114+
"evidence": "No defense pattern found"
115+
}
116+
],
117+
"unicodeIssues": {
118+
"found": false,
119+
"evidence": ""
120+
}
121+
}

‎research/mcp-per-server/fetch.json‎

Lines changed: 121 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,121 @@
1+
{
2+
"grade": "F",
3+
"score": 17,
4+
"coverage": "2/12",
5+
"defended": 2,
6+
"total": 12,
7+
"missing": [
8+
"role-escape",
9+
"data-leakage",
10+
"output-manipulation",
11+
"multilang-bypass",
12+
"unicode-attack",
13+
"indirect-injection",
14+
"social-engineering",
15+
"output-weaponization",
16+
"abuse-prevention",
17+
"input-validation-missing"
18+
],
19+
"checks": [
20+
{
21+
"id": "role-escape",
22+
"name": "Role Boundary",
23+
"nameZh": "角色邊界",
24+
"defended": false,
25+
"confidence": 0.8,
26+
"evidence": "No defense pattern found"
27+
},
28+
{
29+
"id": "instruction-override",
30+
"name": "Instruction Boundary",
31+
"nameZh": "指令邊界",
32+
"defended": true,
33+
"confidence": 0.7,
34+
"evidence": "Found: \"refuse\""
35+
},
36+
{
37+
"id": "data-leakage",
38+
"name": "Data Protection",
39+
"nameZh": "資料保護",
40+
"defended": false,
41+
"confidence": 0.8,
42+
"evidence": "No defense pattern found"
43+
},
44+
{
45+
"id": "output-manipulation",
46+
"name": "Output Control",
47+
"nameZh": "輸出控制",
48+
"defended": false,
49+
"confidence": 0.8,
50+
"evidence": "No defense pattern found"
51+
},
52+
{
53+
"id": "multilang-bypass",
54+
"name": "Multi-language Protection",
55+
"nameZh": "多語言防護",
56+
"defended": false,
57+
"confidence": 0.8,
58+
"evidence": "No defense pattern found"
59+
},
60+
{
61+
"id": "unicode-attack",
62+
"name": "Unicode Protection",
63+
"nameZh": "Unicode 防護",
64+
"defended": false,
65+
"confidence": 0.8,
66+
"evidence": "No defense pattern found"
67+
},
68+
{
69+
"id": "context-overflow",
70+
"name": "Length Limits",
71+
"nameZh": "長度限制",
72+
"defended": true,
73+
"confidence": 0.7,
74+
"evidence": "Found: \"Maximum number of char\""
75+
},
76+
{
77+
"id": "indirect-injection",
78+
"name": "Indirect Injection Protection",
79+
"nameZh": "間接注入防護",
80+
"defended": false,
81+
"confidence": 0.8,
82+
"evidence": "No defense pattern found"
83+
},
84+
{
85+
"id": "social-engineering",
86+
"name": "Social Engineering Defense",
87+
"nameZh": "社交工程防護",
88+
"defended": false,
89+
"confidence": 0.8,
90+
"evidence": "No defense pattern found"
91+
},
92+
{
93+
"id": "output-weaponization",
94+
"name": "Harmful Content Prevention",
95+
"nameZh": "有害內容防護",
96+
"defended": false,
97+
"confidence": 0.8,
98+
"evidence": "No defense pattern found"
99+
},
100+
{
101+
"id": "abuse-prevention",
102+
"name": "Abuse Prevention",
103+
"nameZh": "濫用防護",
104+
"defended": false,
105+
"confidence": 0.8,
106+
"evidence": "No defense pattern found"
107+
},
108+
{
109+
"id": "input-validation-missing",
110+
"name": "Input Validation",
111+
"nameZh": "輸入驗證",
112+
"defended": false,
113+
"confidence": 0.8,
114+
"evidence": "No defense pattern found"
115+
}
116+
],
117+
"unicodeIssues": {
118+
"found": false,
119+
"evidence": ""
120+
}
121+
}

0 commit comments

Comments
 (0)