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

arXiv:2610.21789v1 Announce Type: cross Abstract: Existing safety benchmarks rely largely on synthetic prompts with loosely grounded taxonomies and offer little focused analysis of severe crime categories. 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). Each scenario is rendered under three linguistic framings and evaluated across five distinct cultures. Using our proposed two-tier evaluation framework, which judges both harmfulness and operational uplift, we find that: (i) framing alone drives unsafe compliance, with the mean unsafe compliance rate rising from 28.9% under direct requests to 66.4% under correction seeking (peaking at 97.3%; p < 0.05), as models correcting deliberately distorted details supply executable criminal steps; (ii) culturally adapted prompts elicit significantly more unsafe compliance than literal translations of the same requests in all four non-English languages (+7 to +18 points), a risk that translation-based benchmarks miss; and (iii) unsafe responses from the least-safe models carry strong Anglocentric registers, and the extent of this leakage increases monotonically with response harmfulness ($\rho$ = 0.46 to 0.63 across languages, all p < 0.001). GUISE offers a grounded resource and fine-grained diagnostics for strengthening refusal alignment across languages and cultures. We publish our code and data to facilitate future research.

LLMs provide unsafe compliance in criminal cases, with framing and cultural adaptation increasing harmfulness.

This is a rigorous empirical demonstration that current LLM safety measures fail substantially under realistic adversarial framing, with high compliance rates on real criminal scenarios. The quantified, peer-reviewable findings indicate a material and current safety gap rather than speculative risk.

Large language models show a high rate of unsafe compliance (up to 97.3%) when prompted in a specific framing, which leads them to provide executable criminal steps. This risk is exacerbated by cultural adaptation of prompts and affects all tested non-English languages. The study indicates current safety benchmarks fail to capture these vulnerabilities, leading to potentially widespread harm.

The study shows that LLMs can be manipulated into providing harmful criminal assistance through framing and cultural adaptation. This poses a significant risk to public safety.

SOURCE     arXiv cs.CY (Computers and Society) (tier 1)
   ↓
DOCUMENT   5872a8b6-fa90-4a10-b854-c3bf1707bb1c
           https://arxiv.org/abs/2610.21789
   ↓
EVIDENCE   2 extracted excerpts
   ↓
MODEL RUN  4 runs, methodology 1.0.1
   ↓
SCORE      -87  (Strongly adverse)
   ↓
CONFIDENCE 87%