Download README.md from masahiroid/injecagent-ja: direct link, hf CLI and curl.
- Browser
- Download file 6.7 kB
-
https://huggingface.co/datasets/masahiroid/injecagent-ja/resolve/main/README.md
- Command line
-
hf download hf://datasets/masahiroid/injecagent-ja/README.md
-
curl -L -o README.md https://huggingface.co/datasets/masahiroid/injecagent-ja/resolve/main/README.md
license: mit
language:
- ja
- en
tags:
- prompt-injection
- indirect-prompt-injection
- llm-agents
- tool-use
- llm-security
- japanese
- translation
size_categories:
- 1K<n<10K
pretty_name: InjecAgent-ja
InjecAgent-ja (v0.1.1, translation-first)
Summary
A Japanese translation of InjecAgent (Zhan et al., ACL Findings 2024; MIT): 1,054 test cases (510 direct-harm + 544
data-stealing) x 2 settings (base, enhanced) for evaluating indirect prompt injection against tool-using LLM agents.
- Translated: attacker instructions, expected achievements, user instructions, the agent's
Thought, and the natural-language values inside tool-response templates; theenhancedhacking prefix.Tool Responseis rebuilt from the translated template with the exact rule that reproduces the English field for all 2,108 original cases (checked inbuild_injecagent_ja.py). - Kept as is (so the original harness still applies): tool names, parameters, IDs, e-mail addresses, dollar amounts, dates.
tools.jsonis the original English tool schema;tools_ja.jsonis its Japanese version (v0.1.1): only the natural-language descriptions (category, summaries, parameter/return/exception descriptions) are translated by Amazon Nova Pro; tool, parameter and exception names, types and required flags are unchanged (checked intranslate_injecagent_tools.py). - Files:
test_cases_{dh,ds}_{base,enhanced}_ja.json,attacker_cases_{dh,ds}_ja.jsonl,user_cases_ja.jsonl,tools.json,tools_ja.json, and the translation sources (ja_translations.py,unique_en.json,build_injecagent_ja.py). - Only 175 unique strings are translated (the 2,108 tool responses are composed from them), so a review can cover everything.
v0.1 is a faithful translation, not a localization: names, currencies and places stay as in the original. Japan-specific variants (keigo / indirect phrasing, full-width and mixed scripts, yen and Japanese services) are planned for later versions. Translated by Claude; not yet reviewed by a native-speaker panel. Please report awkward or wrong lines.
Intended use: measuring and hardening the robustness of LLM agents. The attack strings are the benchmark's original test content, translated; they involve fictional data and tools only.
Related: Japanese indirect prompt-injection probes, agentdojo-ja, model-audit-lite.
License / attribution: MIT, Copyright (c) 2023 Qiusi Zhan (original InjecAgent); this translation is released under the same license. Please cite the original:
@inproceedings{zhan2024injecagent,
title={InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents},
author={Zhan, Qiusi and Liang, Zhixiang and Ying, Zifan and Kang, Daniel},
booktitle={Findings of the Association for Computational Linguistics: ACL 2024},
year={2024},
url={https://arxiv.org/abs/2403.02691}
}
Hub: this dataset is part of japanese-llm-security (which resource answers which question).
Cite this dataset
DOI: 10.57967/hf/10738. If you use this dataset, please cite it together with the original work(s) credited above.
@misc{nakatsugawa2026injecagentja,
author = {Nakatsugawa, Masahiro},
title = {InjecAgent-ja},
year = {2026},
publisher = {Hugging Face},
doi = {10.57967/hf/10738},
url = {https://huggingface.co/datasets/masahiroid/injecagent-ja}
}
概要
InjecAgent(Zhan et al., ACL Findings 2024、MIT)の日本語訳です。ツールを使うLLMエージェントに対する間接プロンプト
インジェクションを評価する1,054件(直接的な害 510 + データ窃取 544)× 2設定(base / enhanced)です。
- 翻訳したもの: 攻撃者の指示、期待される達成目標、ユーザーの指示、エージェントの
Thought、ツール応答テンプレート内の 自然文の値、enhancedの指示上書きの前置き。Tool Responseは、英語版の全2,108件を完全に再現する規則(build_injecagent_ja.pyで検証)で、翻訳済みテンプレートから組み立て直しています。 - そのままにしたもの(元のハーネスがそのまま使えるように): ツール名、パラメータ、ID、メールアドレス、ドル金額、日付。
tools.jsonは元の英語のツール定義、**tools_ja.json** はその日本語版です(v0.1.1)。説明文(カテゴリ、概要、パラメータ・戻り値・例外の説明)だけを Amazon Nova Pro で翻訳し、ツール名・パラメータ名・例外名・型・必須フラグは不変です(translate_injecagent_tools.pyで検証)。 - 翻訳対象のユニークな文字列は175件だけ(2,108件の応答はそこから組み立て)なので、全体をレビューできます。
v0.1は忠実な翻訳であり、ローカライズではありません。人名・通貨・地名は原文のままです。日本固有の派生(敬語・婉曲表現、 全角や表記の混在、円と日本のサービス)は今後の版で追加します。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・誤りの指摘を歓迎します。
用途: LLMエージェントの頑健性の測定と強化。攻撃文は元ベンチマークのテスト内容の翻訳で、架空のデータとツールのみを扱います。
ライセンス・帰属: MIT。原著作権 (c) 2023 Qiusi Zhan(InjecAgent)。本翻訳も同じライセンスです。利用時は上記の論文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。
このデータセットの引用
DOI: 10.57967/hf/10738。このデータセットを使う場合は、上に挙げた元の研究とあわせて引用してください。
@misc{nakatsugawa2026injecagentja,
author = {Nakatsugawa, Masahiro},
title = {InjecAgent-ja},
year = {2026},
publisher = {Hugging Face},
doi = {10.57967/hf/10738},
url = {https://huggingface.co/datasets/masahiroid/injecagent-ja}
}