Japanese LLM / agent security
Collection
Japanese indirect prompt-injection datasets (translation-first). Hub: github.com/masahirocom/japanese-llm-security • 4 items • Updated
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
Information Retrieval: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Content Creation: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Learning and Tutoring: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Language Translation: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Programming Help: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Alphanumeric Substitution: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Homophonic Substitution: struct<category_ja: string, script_dependent: bool, attacks: list<
...
script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Cookie Theft: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Crippling Critical Infrastructures: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Cryptocurrency Mining: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Dumpster Diving: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Device and Driver Enumeration: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Exploiting System Vulnerabilities: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
to
{'Cookie Theft': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Memory Scanning': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Dumpster Diving': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Environment Variable Analysis': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Device and Driver Enumeration': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Sending Out Spam Emails': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Crippling Critical Infrastructures': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Network Propagation': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Exploiting System Vulnerabilities': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Cryptocurrency Mining': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(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 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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
Information Retrieval: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Content Creation: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Learning and Tutoring: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Language Translation: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Programming Help: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Alphanumeric Substitution: struct<category_ja: string, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: string
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Homophonic Substitution: struct<category_ja: string, script_dependent: bool, attacks: list<
...
script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Cookie Theft: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Crippling Critical Infrastructures: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Cryptocurrency Mining: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Dumpster Diving: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Device and Driver Enumeration: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
Exploiting System Vulnerabilities: struct<category_ja: null, script_dependent: bool, attacks: list<item: string>>
child 0, category_ja: null
child 1, script_dependent: bool
child 2, attacks: list<item: string>
child 0, item: string
to
{'Cookie Theft': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Memory Scanning': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Dumpster Diving': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Environment Variable Analysis': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Device and Driver Enumeration': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Sending Out Spam Emails': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Crippling Critical Infrastructures': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Network Propagation': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Exploiting System Vulnerabilities': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}, 'Cryptocurrency Mining': {'category_ja': Value('null'), 'script_dependent': Value('bool'), 'attacks': List(Value('string'))}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
BIPIA-attacks-ja (v0.2)
Summary
A Japanese translation of the attack files of BIPIA (Yi et al., "Benchmarking and Defending Against Indirect Prompt Injection Attacks on Large Language Models"; MIT, Microsoft): 29 text-attack categories and 20 code-attack categories, 5 attacks each, train and test splits (250 attacks in total).
- Translated: the instruction sentences. Kept verbatim: code blocks (byte-for-byte), URLs, e-mail addresses and quoted
payload strings such as
'Hello, how are you?'(validated inbuild_bipia_ja.py). - Each category has
category_jaand ascript_dependentflag. Categories that assume the Latin alphabet (substitution / homophonic / misspelling / anagram / space removal / Caesar / Base-N / reverse text) are translated faithfully but remain alphabet-based; Japanese-native variants (kana, kanji, full-width) are planned for a later version. - Not included: BIPIA's context files (email / table / code / web QA / summarization), which derive from other datasets with their own terms. Build them with the original BIPIA scripts and combine them with these attack files.
- New in v0.2 — Japanese-native attacks:
text_attack_ja_native.json, 10 categories x 5 (50 attacks) that exploit Japanese writing systems and register instead of the Latin alphabet: hiragana / katakana / half-width kana / full-width alphanumerics / romanization, a gojuon-shift cipher, reversed Japanese text, homophone-kanji substitution, register and keigo changes (casual, over-polite, Kansai dialect), punctuation and spacing manipulation. They are the Japanese counterparts of thescript_dependentcategories above (original to this project; MIT). Instruction sentences only. - Files:
text_attack_{train,test}_ja.json,code_attack_{train,test}_ja.json, plusbipia_ja_translations.pyandbuild_bipia_ja.py(to rebuild from the original repository).
Translated by Claude; not yet reviewed by a native-speaker panel. Please report awkward lines. Related: InjecAgent-ja, Japanese indirect prompt-injection probes, agentdojo-ja.
Intended use: measuring and hardening the robustness of LLM applications against indirect prompt injection. The attack strings are the benchmark's original test content, translated.
License / attribution: MIT, Copyright (c) Microsoft Corporation (original BIPIA); this translation is under the same license. Please cite the original paper (arXiv:2312.14197).
Hub: this dataset is part of japanese-llm-security (which resource answers which question).
概要
BIPIA（Yi et al.「Benchmarking and Defending Against Indirect Prompt Injection Attacks on Large Language Models」、 MIT、Microsoft）の攻撃ファイルの日本語訳です。テキスト攻撃29カテゴリ、コード攻撃20カテゴリ、各5件、train/test、計250件。
- 翻訳: 指示文。そのまま保持: コードブロック（バイト単位）、URL、メールアドレス、
'Hello, how are you?'のような引用された ペイロード文字列（build_bipia_ja.pyで検証）。 - 各カテゴリに
category_jaとscript_dependent（ラテン文字を前提とするカテゴリ）を付けています。置換・同音・綴り間違い・アナグラム・ 空白除去・シーザー暗号・Base系・反転は忠実に訳したうえでアルファベット前提のままです。日本語固有の版（かな・漢字・全角）は今後追加します。 - v0.2の新規 — 日本語固有の攻撃:
text_attack_ja_native.json（10カテゴリ×5件＝50件）。ラテン文字ではなく日本語の表記体系と文体を突くもの: ひらがな化・カタカナ化・半角カナ化・全角英数化・ローマ字化、五十音ずらし暗号、日本語文字列の反転、同音異字、文体・敬語レベルの変更（タメ口・過剰敬語・関西弁）、句読点・空白の操作。 上のscript_dependentカテゴリの日本語版です（本プロジェクトの独自作成、MIT）。指示文のみです。 - 含まないもの: BIPIAの文脈データ（メール・表・コード・Web QA・要約）は、他のデータセット由来で条件が異なるため含めていません。 元のBIPIAのスクリプトで作成し、本データの攻撃文と組み合わせてください。
- 翻訳はClaudeによるもので、ネイティブによる査読は未実施です。不自然な箇所の指摘を歓迎します。
ライセンス・帰属: MIT。原著作権 (c) Microsoft Corporation（BIPIA）。本翻訳も同じライセンスです。原論文（arXiv:2312.14197）を引用してください。
ハブ: このデータセットは japanese-llm-security（どのリソースが何の問いに答えるか）の一部です。
- Downloads last month
- 44