Global Security Blind Spots: LLM Safety Failures in Low-Resource Languages

In Summary

Welo Data’s latest research examines how large language model (LLM) safety degrades beyond English. 

By evaluating harmful prompts across 79 languages under controlled conditions, the study shows that models which appear well aligned in English frequently produce unsafe responses in low-resource languages. In many cases, simple translation alone is enough to bypass existing safety guardrails. 

The findings reveal a structural weakness in current safety alignment approaches and highlight multilingual safety as a critical and undermeasured global security risk since weaknesses in any language can be exploited to bypass protections across the entire system.  

What you’ll learn

• How unsafe response rates change when prompts move from English to low-resource languages 
• Why translation can function as a practical jailbreak for modern LLMs 
• Which language families exhibit the highest safety risk 
• Which harm categories show the largest cross-lingual safety gaps 
• How English-only safety benchmarks can misrepresent real-world deployment risk