Datasets:

promptprotection
/

agent-security-datasets

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
id: string
label: int64
category: string
technique: string
text: string
call: null
sinks: null
user: string
scenario: string
expect: string
notes: string
expect_reason: string
known_miss: string
sources: list<item: struct<id: string, tool: string, text: string>>
  child 0, item: struct<id: string, tool: string, text: string>
      child 0, id: string
      child 1, tool: string
      child 2, text: string
to
{'id': Value('string'), 'label': Value('string'), 'scenario': Value('string'), 'user': Value('string'), 'sources': List({'id': Value('string'), 'tool': Value('string'), 'text': Value('string')}), 'call': {'toolName': Value('string'), 'args': Json(decode=True)}, 'sinks': Json(decode=True), 'expect': Value('string'), 'expect_reason': Value('string'), 'notes': Value('string'), 'known_miss': Value('string')}
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
              id: string
              label: int64
              category: string
              technique: string
              text: string
              call: null
              sinks: null
              user: string
              scenario: string
              expect: string
              notes: string
              expect_reason: string
              known_miss: string
              sources: list<item: struct<id: string, tool: string, text: string>>
                child 0, item: struct<id: string, tool: string, text: string>
                    child 0, id: string
                    child 1, tool: string
                    child 2, text: string
              to
              {'id': Value('string'), 'label': Value('string'), 'scenario': Value('string'), 'user': Value('string'), 'sources': List({'id': Value('string'), 'tool': Value('string'), 'text': Value('string')}), 'call': {'toolName': Value('string'), 'args': Json(decode=True)}, 'sinks': Json(decode=True), 'expect': Value('string'), 'expect_reason': Value('string'), 'notes': Value('string'), 'known_miss': Value('string')}
              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.

prompt-protection datasets

A held-out, human-authored corpus for evaluating prompt-injection / agent-security guards. Every item is original to this project and disjoint from the unit-test fixtures (tests/__fixtures__/*.txt) and bench corpus (bench/corpus/*), so it measures generalisation rather than memorisation. JSONL, one object per line, UTF-8.

Files

`agent-flows.jsonl`, tool-call guard scenarios (100: 50 attack / 50 benign)

Each row is a full agent turn: the user's instruction, the tool results (sources, which may carry an injection), the tool call the model wants to make, an explicit sinks map (tool → network|email|message|file-write|exec|payment|none), and the expected guard decision. Fields: id, label, scenario, user, sources[], call{toolName,args}, sinks, expect (block|flag|allow), expect_reason, notes.

`benign-hard.jsonl`, over-defense discipline (155, all `label:0`)

Legitimate prompts that contain injection trigger vocabulary ("ignore", "override", "jailbreak", "system prompt"). Categories: common-query, technique-query, virtual-creation, multilingual, dev-jargon, security-docs. A guard that blocks these is over-defending. Fields: id, label, category, text, triggers[].

`attacks.jsonl`, regex-evading attacks (130, all `label:1`)

Attacks a keyword scanner tends to miss: paraphrased overrides with no canonical phrase, persona jailbreaks with no trigger words, encoding/obfuscation (base64, homoglyph, zero-width, ROT13, hex, spaced), indirect injections embedded in realistic emails/web/README/CSV, markdown-image/link exfil, role-tag forgery, MCP tool poisoning, and 15 multilingual items across 14 languages. Fields: id, label, category, technique, text.

How labels & expectations were assigned

attacks / benign-hard labels are definitional (malicious intent = 1, legitimate = 0).
agent-flows.expect follows the guard's documented policy order (block > confirm > flag): a tainted identifier or ≥12-char verbatim payload reaching an exfil/exec sink → block; an injection-scoring source with a same-turn sink but no shared identifier → flag (injection-source-then-sink); any payment sink, and cases where the recipient/URL comes from a tool result rather than being named by the user, → flag with confirm semantics (documented in notes); user-named recipients/URLs and read-only sinks → allow. The ~5 benign-labelled rows that expect flag are intentional confirm cases, each noted.

Disjointness

validate.mjs normalises every text and asserts zero overlap with the four fixture files. Run node datasets/validate.mjs (Node ≥20, no deps) to re-check schema, unique ids, enum values, per-category counts, and disjointness.

Licence & citation

Licensed CC-BY-4.0 (see LICENSE). If you use this corpus, please cite:

prompt-protection agent-security datasets (2026), https://github.com/mughalhere/prompt-protection, CC-BY-4.0.

Contributing

Add rows that keep each file's schema and stay disjoint from the fixtures; prefer techniques a pure-regex scanner would miss for attacks, and realistic legitimate uses of trigger vocabulary for benign-hard. Run the validator before opening a PR, it must print OK.

Benchmark snapshot (from bench/results.json)

Set	N	Recall (regex)	FP rate (regex)
local-tuning	134	100.0%	0.0%
local-heldout	35	75.0%	6.7%
datasets/attacks+benign-hard	285	14.6%	19.4%
NotInject (over-defence)	339	n/a	2.9%
in-the-wild (sample)	900	46.3%	20.5%
agent-flows (guard)	100	block-recall 82.0%	benign FPR 4.0%

Downloads last month: 97

Total file size:

139 kB