Welo Data researchers found that harmful prompts handled safely in English produced unsafe responses when translated into low-resource languages, with average bypass rates ranging from 21.9% to 44.4% across ten language models. The findings, published in Global Security Blind Spots: Translation as a Safety Bypass in Low-Resource Languages, highlight significant gaps in English-only AI safety testing.
The study tested 119 harmful prompts translated from English into 78 low-resource languages spanning 20 language families. The prompts covered 11 safety categories, including dangerous behavior and criminal content, violence and threats, hate and discrimination, self-harm and suicide, and misinformation. By measuring unsafe responses only when models handled the original English prompt safely, researchers distinguished translation-based safety failures from unsafe behavior already present in English. The researchers also found that bypass patterns varied across language families and differed between models, highlighting the importance of evaluating multilingual safety beyond aggregate scores.
"AI models that resist harmful requests in English may be much more vulnerable to the same requests in other languages. That creates opportunities for adversaries to exploit safety gaps that English-only testing would never reveal," said David Harper, Data Scientist at Welo Data and co-author of the paper.
English-only safety testing cannot establish whether a model’s protections extend to the other languages it supports, leaving multilingual users exposed to risks that conventional evaluations may miss. The paper recommends tracking translation-based bypass rates alongside overall unsafe response rates and auditing performance across language families. For model builders, the findings underscore the importance of multilingual safety evaluations that incorporate native-language expertise, adversarial testing, and human validation.
“Building a model that works across languages is only half the challenge. Ensuring it behaves safely across those languages is just as important. Multilingual safety evaluation needs to be treated as a core requirement for global AI deployment, not an afterthought,” said Siobhan Hanna, SVP and GM, Welo Data.
The paper is available on arXiv. Findings from the study will be presented at NeurIPS 2026 in Paris.
Welo Data, a Welo Global brand, is the data production and evaluation partner for foundation model builders and enterprises deploying AI at scale. Through a global network of 500,000+ vetted experts across 300+ languages and locales, Welo Data delivers the human judgment, structured evaluation, and data infrastructure that AI programs require to perform reliably across languages, cultures, and real-world contexts. Programs span large language models, voice and speech systems, agentic workflows, and robotics. Welo Data’s proprietary identity and fraud-prevention framework, NIMO, maintains data integrity and workforce accountability across every engagement.
