Global Security Blind Spots: Translation As a Safety Bypass in Low-Resource Languages

Large language models are deployed globally, yet safety alignment and evaluation remain concentrated in English: harmful requests refused in English can elicit unsafe completions when translated into low-resource languages. We evaluate ten widely deployed LLMs across 79 languages English plus 78...

MAI-2023-0007

Large Language Models LLMs, including GPT-4, are susceptible to cross-lingual vulnerabilities within their safety mechanisms. By translating unsafe English prompts into low-resource languages using widely accessible translation services such as Google Translate, attackers can bypass the LLM's...

MAI-2023-0008

Large Language Models LLMs, including GPT-4, are susceptible to a cross-lingual vulnerability that compromises their safety mechanisms. This vulnerability is exploited by translating unsafe prompts from English into low-resource languages using widely accessible translation services such as Googl...