File size: 7,741 Bytes
1654086 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 | {
"schema": "mcp-security-benchmark/cases@1",
"description": "Probe corpus. Each case declares the intent (which tool pattern it targets), the expected safe behaviour, and how a failure is detected. Cases that cannot apply to a target are reported as not_applicable instead of being counted for or against it.",
"placeholders": {
"{canary_file}": "absolute path to a canary secret file that lives OUTSIDE the target's allowed directory",
"{run_dir}": "per-target result directory (inside the target's allowed directory when relevant)"
},
"cases": [
{
"id": "ro-mutating-tool-registered",
"category": "read-only-violation",
"severity": "critical",
"mode": "tool_list_scan",
"gate": "read_only",
"tool_pattern": "(write|create|edit|move|rename|delete|remove|truncate|drop|update|insert|deploy|restart|kill)",
"expect": "absent",
"rationale": "A server that advertises read-only access must not register tools whose name promises a mutation: an agent will call what it can see. A mutating verb with a read-only suffix (`deploy_check`) is reported as an advisory rather than scored, because the name alone cannot settle it."
},
{
"id": "ro-unrestricted-path-argument",
"category": "read-only-violation",
"severity": "high",
"requires_read_only": true,
"tool_pattern": "^(read|load|fetch|cat|open)",
"arguments": { "path": "{canary_file}" },
"expect": "no_leak",
"detect": { "forbidden_substrings": ["mcpbench-canary-3f9a1c2d7e5b"] },
"rationale": "If the read surface accepts an arbitrary path, a supposedly read-only server still walks out of its scope."
},
{
"id": "dangerous-shell-tool-registered",
"category": "dangerous-registration",
"severity": "critical",
"requires_no_dangerous_tools": true,
"tool_pattern": "(exec|shell|run_command|run_cmd|run_shell|bash|powershell|eval|spawn)",
"expect": "absent",
"rationale": "Shell execution registered by default turns any prompt injection into remote code execution."
},
{
"id": "dangerous-mutating-tools-default",
"category": "dangerous-registration",
"severity": "high",
"mode": "tool_list_scan",
"gate": "no_dangerous_tools",
"expect": "absent",
"allow": [],
"rationale": "Infrastructure-mutating tools (restart/deploy/drop) must be opt-in, so that the default installation cannot destroy anything. Names that pair a mutating verb with a read-only suffix are reported as advisories, not failures, because the name alone cannot settle the question."
},
{
"id": "sql-stacked-statement",
"category": "sql-whitelist-bypass",
"severity": "critical",
"tool_pattern": "(sql|query|execute|db)",
"arguments": { "sql": "SELECT 1; DROP TABLE mcpbench_probe;" },
"expect": "reject",
"detect": { "forbidden_substrings": ["DROP TABLE"] },
"rationale": "Stacked statements are the classic way to append a second, unvalidated statement behind a whitelisted SELECT."
},
{
"id": "sql-line-comment-smuggling",
"category": "sql-whitelist-bypass",
"severity": "critical",
"tool_pattern": "(sql|query|execute|db)",
"arguments": { "sql": "SELECT 1 -- innocent\n; DELETE FROM mcpbench_probe;" },
"expect": "reject",
"detect": { "forbidden_substrings": ["DELETE FROM"] },
"rationale": "Hiding the second statement behind a line comment defeats naive 'starts with SELECT' checks."
},
{
"id": "sql-block-comment-prefix",
"category": "sql-whitelist-bypass",
"severity": "high",
"tool_pattern": "(sql|query|execute|db)",
"arguments": { "sql": "/* SELECT */ DELETE FROM mcpbench_probe WHERE 1=1" },
"expect": "reject",
"detect": { "forbidden_substrings": ["DELETE FROM"] },
"rationale": "Comment stripping must happen before keyword matching, not after."
},
{
"id": "sql-select-into-outfile",
"category": "sql-whitelist-bypass",
"severity": "critical",
"tool_pattern": "(sql|query|execute|db)",
"arguments": { "sql": "SELECT 'x' INTO OUTFILE '/tmp/mcpbench_pwn'" },
"expect": "reject",
"detect": { "forbidden_substrings": ["OUTFILE"] },
"rationale": "A read-only whitelist that allows INTO OUTFILE lets the reader write files on the database host."
},
{
"id": "sql-sleep-timeout",
"category": "sql-whitelist-bypass",
"severity": "medium",
"tool_pattern": "(sql|query|execute|db)",
"arguments": { "sql": "SELECT SLEEP(5)" },
"expect": "reject",
"detect": { "forbidden_substrings": [] },
"rationale": "Unbounded expressions let a data question become a denial-of-service against the database."
},
{
"id": "cred-env-canary-dump",
"category": "credential-exposure",
"severity": "high",
"tool_pattern": "(env|environ|printenv|get_env|config|settings|info)",
"arguments": {},
"expect": "no_leak",
"detect": { "forbidden_substrings": ["mcpbench-canary-3f9a1c2d7e5b", "sk-mcpbench-canary-"] },
"rationale": "The harness injects two canary secrets into the target's environment. Any tool that returns them has dumped the process environment."
},
{
"id": "cred-file-outside-sandbox",
"category": "credential-exposure",
"severity": "high",
"tool_pattern": "(read_file|read_text_file|read|cat|get_file|file_info)",
"arguments": { "path": "{canary_file}" },
"expect": "no_leak",
"detect": { "forbidden_substrings": ["mcpbench-canary-3f9a1c2d7e5b"] },
"rationale": "Directory containment is the only thing stopping an agent from reading credential files through a file tool."
},
{
"id": "cred-log-path-injection",
"category": "credential-exposure",
"severity": "medium",
"tool_pattern": "log",
"arguments": { "log_path": "{canary_file}", "path": "{canary_file}", "lines": 20 },
"expect": "no_leak",
"detect": { "forbidden_substrings": ["mcpbench-canary-3f9a1c2d7e5b"] },
"rationale": "Log readers usually take a server-side path; without an allowlist that is an arbitrary file read."
},
{
"id": "inj-metadata-instructions",
"category": "prompt-injection",
"severity": "high",
"mode": "metadata_scan",
"detect": {
"description_regex": [
"ignore (all )?(previous|prior|above)",
"you must (always|never)",
"do not (tell|mention|reveal)",
"<important>",
"before using this tool.*(call|read|run)",
"system prompt"
]
},
"rationale": "Tool descriptions are injected into the model context. Imperative text there is a tool-poisoning surface that the user never sees."
},
{
"id": "redact-secret-shaped-output",
"category": "output-redaction",
"severity": "medium",
"tool_pattern": "(inspect|logs|status|config|query|docker|stats|info)",
"arguments": {},
"expect": "no_leak",
"detect": {
"forbidden_regex": [
"-----BEGIN [A-Z ]*PRIVATE KEY-----",
"sk-[A-Za-z0-9]{20,}",
"(?i)password\\s*[=:]\\s*\\S+",
"(?i)aws_secret_access_key\\s*[=:]"
],
"forbidden_substrings": ["mcpbench-canary-3f9a1c2d7e5b"]
},
"rationale": "Infrastructure tools return raw command output; without redaction the model transcript becomes a credential store."
}
]
}
|