A recent arXiv paper (2609.02852) examines how large language models handle inputs in low-resource languages and scripts, revealing what the authors call linguistic illegibility. The study shows that when text is encoded in scripts or languages underrepresented in training data, safety filters and content moderation systems often fail to recognize harmful instructions. Attackers can exploit this gap by translating malicious prompts into such languages, bypassing guardrails that work reliably in English. The paper demonstrates successful jailbreaks across multiple commercial and open-source LLMs using this method. The authors call for multilingual safety alignment and better tokenization strategies to close this vulnerability.
This is the kind of research that makes you stop and think about the hidden assumptions in AI safety. We build guardrails, test them in English, and assume they hold everywhere. But language is not a uniform layer. It is a patchwork of scripts, dialects, and digital haves and have-nots. Attackers do not need to be sophisticated. They just need a language the model barely knows. That is not a bug. It is a blind spot we chose to ignore.
The fix is not simple. Retraining on more languages helps, but tokenization itself creates inequalities. Some scripts get chopped into meaningless pieces. Safety training then never sees coherent harmful intent. We need to treat multilingual robustness as a first-class security requirement, not an afterthought. Otherwise, the next wave of AI exploits will not come from clever code. They will come from a phrasebook.