In the Guise of Correction: How Framing and Cultural Adaptation Erode LLM Safety on Real-World Criminal Cases

Key points

• Existing safety benchmarks rely largely on synthetic prompts with loosely grounded taxonomies and offer little focused analysis of severe crime categories.1
• Each scenario is rendered under three linguistic framings and evaluated across five distinct cultures.1
• GUISE offers a grounded resource and fine-grained diagnostics for strengthening refusal alignment across languages and cultures.1
• We publish our code and data to facilitate future research.1

Sources (1)

• [1]In the Guise of Correction: How Framing and Cultural Adaptation Erode LLM Safety on Real-World Criminal Cases
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 12:44 PM

  We introduce GUISE, a benchmark of criminal-assistance scenarios sourced from real-world cases and organised under a taxonomy from the UNODC International Classification of Crime for Statistical Purposes (ICCS).

  Existing safety benchmarks rely largely on synthetic prompts with loosely grounded taxonomies and offer little focused analysis of severe crime categories.

Extractive summary: sentences quoted from the sources.