The authors adopt a two-step threat model: first probe a model with a harmful prompt in English; if refused, translate the same prompt into a low-resource language and retry.
Translation bypasses safety alignment in low-resource languages across multiple LLMs
Authors evaluate 10 widely deployed LLMs across 79 languages, finding mean bypass rates of 22-44% when English responses are safe and no stable model ranking.
Industry
Jalil Alva · David Harper · Abigail E. Thornton · Rafael S. Villalobos Alva · Leonardo Kenji Minemura Suazo · Konstantinos Karageorgos · +5 more
Welo Data
Research Digest··1 min read
The authors test whether English safety alignment transfers to low-resource languages by translating harmful prompts that models refuse in English.
§