Datasets:

CieveHe
/

mcp-security-benchmark

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
dataset: string
version: string
generated: timestamp[s]
source_repository: string
doi: string
concept_doi: string
note: string
targets: struct<everything-public: struct<target: string, target_file: string, notes: string, declared_read_o (... 11173 chars omitted)
  child 0, everything-public: struct<target: string, target_file: string, notes: string, declared_read_only: bool, expect_no_dange (... 2343 chars omitted)
      child 0, target: string
      child 1, target_file: string
      child 2, notes: string
      child 3, declared_read_only: bool
      child 4, expect_no_dangerous_tools: bool
      child 5, handshake: struct<ok: bool, protocol_version: timestamp[s], server_info: struct<name: string, title: string, ve (... 434 chars omitted)
          child 0, ok: bool
          child 1, protocol_version: timestamp[s]
          child 2, server_info: struct<name: string, title: string, version: string>
              child 0, name: string
              child 1, title: string
              child 2, version: string
          child 3, raw: struct<protocolVersion: timestamp[s], capabilities: struct<tools: struct<listChanged: bool>, prompts (... 312 chars omitted)
              child 0, protocolVersion: timestamp[s]
              child 1, capabilities: struct<tools: struct<listChanged: bool>, prompts: struct<listChanged: bool>, resources: struct<subsc (... 171 chars omitted)
                  child 0, tools: struct<listChanged: bool>
                      child 0, listChanged: bool
 
...
ct: string
cases: list<item: struct<id: string, category: string, severity: string, mode: string, gate: string, tool_p (... 349 chars omitted)
  child 0, item: struct<id: string, category: string, severity: string, mode: string, gate: string, tool_pattern: str (... 337 chars omitted)
      child 0, id: string
      child 1, category: string
      child 2, severity: string
      child 3, mode: string
      child 4, gate: string
      child 5, tool_pattern: string
      child 6, expect: string
      child 7, rationale: string
      child 8, requires_read_only: bool
      child 9, arguments: struct<path: string, sql: string, log_path: string, lines: int64>
          child 0, path: string
          child 1, sql: string
          child 2, log_path: string
          child 3, lines: int64
      child 10, detect: struct<forbidden_substrings: list<item: string>, description_regex: list<item: string>, forbidden_re (... 24 chars omitted)
          child 0, forbidden_substrings: list<item: string>
              child 0, item: string
          child 1, description_regex: list<item: string>
              child 0, item: string
          child 2, forbidden_regex: list<item: string>
              child 0, item: string
      child 11, requires_no_dangerous_tools: bool
      child 12, allow: list<item: null>
          child 0, item: null
placeholders: struct<{canary_file}: string, {run_dir}: string>
  child 0, {canary_file}: string
  child 1, {run_dir}: string
description: string
schema: string
to
{'schema': Value('string'), 'description': Value('string'), 'placeholders': {'{canary_file}': Value('string'), '{run_dir}': Value('string')}, 'cases': List({'id': Value('string'), 'category': Value('string'), 'severity': Value('string'), 'mode': Value('string'), 'gate': Value('string'), 'tool_pattern': Value('string'), 'expect': Value('string'), 'rationale': Value('string'), 'requires_read_only': Value('bool'), 'arguments': {'path': Value('string'), 'sql': Value('string'), 'log_path': Value('string'), 'lines': Value('int64')}, 'detect': {'forbidden_substrings': List(Value('string')), 'description_regex': List(Value('string')), 'forbidden_regex': List(Value('string'))}, 'requires_no_dangerous_tools': Value('bool'), 'allow': List(Value('null'))})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              dataset: string
              version: string
              generated: timestamp[s]
              source_repository: string
              doi: string
              concept_doi: string
              note: string
              targets: struct<everything-public: struct<target: string, target_file: string, notes: string, declared_read_o (... 11173 chars omitted)
                child 0, everything-public: struct<target: string, target_file: string, notes: string, declared_read_only: bool, expect_no_dange (... 2343 chars omitted)
                    child 0, target: string
                    child 1, target_file: string
                    child 2, notes: string
                    child 3, declared_read_only: bool
                    child 4, expect_no_dangerous_tools: bool
                    child 5, handshake: struct<ok: bool, protocol_version: timestamp[s], server_info: struct<name: string, title: string, ve (... 434 chars omitted)
                        child 0, ok: bool
                        child 1, protocol_version: timestamp[s]
                        child 2, server_info: struct<name: string, title: string, version: string>
                            child 0, name: string
                            child 1, title: string
                            child 2, version: string
                        child 3, raw: struct<protocolVersion: timestamp[s], capabilities: struct<tools: struct<listChanged: bool>, prompts (... 312 chars omitted)
                            child 0, protocolVersion: timestamp[s]
                            child 1, capabilities: struct<tools: struct<listChanged: bool>, prompts: struct<listChanged: bool>, resources: struct<subsc (... 171 chars omitted)
                                child 0, tools: struct<listChanged: bool>
                                    child 0, listChanged: bool
               
              ...
              ct: string
              cases: list<item: struct<id: string, category: string, severity: string, mode: string, gate: string, tool_p (... 349 chars omitted)
                child 0, item: struct<id: string, category: string, severity: string, mode: string, gate: string, tool_pattern: str (... 337 chars omitted)
                    child 0, id: string
                    child 1, category: string
                    child 2, severity: string
                    child 3, mode: string
                    child 4, gate: string
                    child 5, tool_pattern: string
                    child 6, expect: string
                    child 7, rationale: string
                    child 8, requires_read_only: bool
                    child 9, arguments: struct<path: string, sql: string, log_path: string, lines: int64>
                        child 0, path: string
                        child 1, sql: string
                        child 2, log_path: string
                        child 3, lines: int64
                    child 10, detect: struct<forbidden_substrings: list<item: string>, description_regex: list<item: string>, forbidden_re (... 24 chars omitted)
                        child 0, forbidden_substrings: list<item: string>
                            child 0, item: string
                        child 1, description_regex: list<item: string>
                            child 0, item: string
                        child 2, forbidden_regex: list<item: string>
                            child 0, item: string
                    child 11, requires_no_dangerous_tools: bool
                    child 12, allow: list<item: null>
                        child 0, item: null
              placeholders: struct<{canary_file}: string, {run_dir}: string>
                child 0, {canary_file}: string
                child 1, {run_dir}: string
              description: string
              schema: string
              to
              {'schema': Value('string'), 'description': Value('string'), 'placeholders': {'{canary_file}': Value('string'), '{run_dir}': Value('string')}, 'cases': List({'id': Value('string'), 'category': Value('string'), 'severity': Value('string'), 'mode': Value('string'), 'gate': Value('string'), 'tool_pattern': Value('string'), 'expect': Value('string'), 'rationale': Value('string'), 'requires_read_only': Value('bool'), 'arguments': {'path': Value('string'), 'sql': Value('string'), 'log_path': Value('string'), 'lines': Value('int64')}, 'detect': {'forbidden_substrings': List(Value('string')), 'description_regex': List(Value('string')), 'forbidden_regex': List(Value('string'))}, 'requires_no_dangerous_tools': Value('bool'), 'allow': List(Value('null'))})}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

MCP security benchmark v0.1 — probe corpus and scorecards

This dataset holds the probe corpus and the per-target scorecards produced by mcp-security-benchmark, a dependency-free harness that probes Model Context Protocol (MCP) servers over stdio and scores them against a fixed corpus of security-relevant cases.

It exists because "our MCP server is read-only and safe" is a claim, not evidence. The claim becomes evidence when a fixed corpus runs against the server, produces a scorecard, and the harness has been shown to fail a deliberately broken server.

Files

file	contents
`cases.json`	the probe corpus: six threat categories, 14 cases, each with severity and the tool pattern it targets
`scorecards.json`	one structured scorecard per target, including the MCP handshake metadata, the tool surface and every probe result
`summary.md`	the human-readable comparison table for the five targets below

Targets (2026-09-28, one machine)

target	what it is	verdict	score	findings
`vulnerable-demo`	bundled canary target, insecure on purpose	critical	4.8	12
`everything-public`	npm `@modelcontextprotocol/server-everything`	elevated	88.0	1
`mysql-ops-mcp-reference`	read-only-first MySQL ops server (reference implementation)	no findings	100.0	0 (+2 advisories)
`filesystem-public`	npm `@modelcontextprotocol/server-filesystem`	no findings	100.0	0
`memory-public`	npm `@modelcontextprotocol/server-memory`	no findings	100.0	0

The one high finding: everything-public exposes a get-env tool that returns the process environment; the probe recovered both canary secrets the harness injected. For an agent host, any API key in that server's environment is one tool call away. Reported upstream as modelcontextprotocol/servers#4882.

What the score does and does not mean

score = severity-weighted pass rate over applicable probes. A probe whose tool pattern does not exist on a target is not_applicable and moves neither way.
A score of 100 is not "secure". The corpus has 14 cases; it does not cover authentication, transport, sandboxing, dependency risk or the target's own business logic.
Name-based registration checks are heuristics and are reported as advisories that do not change the score (mysql-ops-mcp-reference carries two).
Refusal detection reads the response body, not only isError, because several implementations (FastMCP included) return refusals as ordinary content.
Comparing servers with different declared policies is only meaningful per category; the per-target scorecards carry that detail.
These numbers describe one machine and one corpus version. Re-running the harness on another machine may differ in timing, not in the recorded probe outcomes.

Citation

Cite the archived release:

He, Zhen. (2026). MCP security benchmark: a fixed probe corpus and scorecards for MCP servers (v0.1.0). Zenodo. https://doi.org/10.5281/zenodo.23003717

All versions: https://doi.org/10.5281/zenodo.23003715

Downloads last month: 49

Total file size:

79.3 kB