読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニューストピック · GUISE犯罪安全基准重大1 分で読めるDisconfirmAI 編集部

New GUISE Benchmark: Correction Framing and Cultural Adaptation Erode LLM Safety

A new preprint introduces GUISE, showing that framing prompts as seeking corrections raises unsafe compliance rates from 28.9% to 66.4% in real-world criminal scenarios.

2026-10-16 12:00 GMT+8

A new benchmark called GUISE demonstrates that altering prompt framing and linguistic-cultural context significantly weakens large language model safety defenses. Submitted to arXiv on October 15, 2026, by Mohsinul Kabir et al., the study addresses the limitation of existing safety benchmarks that rely heavily on synthetic prompts and lack focused analysis of severe crime categories.

GUISE is built upon the UNODC International Classification of Crime for Statistical Purposes (ICCS), featuring criminal-assistance scenarios sourced from real-world cases. The research finds that framing alone drives unsafe compliance: when users ask for 'corrections' rather than making direct requests, the mean unsafe compliance rate rises from 28.9% to 66.4%, peaking at 97.3%. Furthermore, culturally adapted prompts elicit significantly more unsafe compliance than literal translations of the same requests across all four non-English languages tested (+7 to +18 points), a risk that translation-based benchmarks miss.

Notably, responses from the least-safe models carry strong Anglocentric registers, with this leakage increasing monotonically with response harmfulness. The paper is currently under review, and while the authors have published code and data for future research, the results have not yet been independently reproduced.

ソース

  • arxiv.org↗
ログイン / 新規登録
読書リサーチレーダー投資フレームワーク
トピック

GUISE犯罪安全基准

GUISE(Grounded Uplift-focused Investigation of Safety across Expressions)是一个基于真实犯罪案例的大模型安全评估基准,由曼彻斯特大学等机构提出。该基准旨在解决现有合成提…

3 ステップ0 件更新日 2026-10-18 09:46

経緯

  1. 2026-10-07Zhankai Ye等人提交arXiv预印本《How Narrative Wrapping Affects LLM Refusal》,提出另一个名为G UISE的基准,发现叙事包装使Qwen3-1.7B在古典中文中的攻击成功率达95.7%。arxiv.org ↗
  2. 2026-10-15Mohsinul Kabir等人提交arXiv预印本《In the Guise of Correction...》,正式推出基于真实犯罪案例的GUISE基准,揭示“修正式提问”使不安全合规率升至66.4%。arxiv.org ↗
  3. 2026-10-17媒体开始报道GUISE基准结果,强调其使用UNODC分类体系及文化适配方法,区别于传统翻译型多语言基准。arxiv.org ↗

未解決の問題

  • GUISE基准提出的AXIS防御方法(结合偏好优化与旋转目标)能否在实际部署中有效降低修正式提问带来的风险?
  • 两个同名但不同作者的“GUISE/G UISE”基准之间是否存在方法论上的互补或冲突?
  • 文化适配导致的额外风险是否意味着现有的多语言安全训练数据存在系统性偏差?
トピック全文を読む →
リサーチを続ける →