Hackers can exploit low-resource languages to jailbreak AI chatbots

LLMs remain vulnerable to jailbreak attacks even when safety training focuses on English. Researchers tested whether conversations in low-resource African languages (Afrikaans, Kiswahili, isiXhosa, isiZulu) could circumvent safety guardrails across five major commercial models. Multi-turn conversations achieved harmful response rates of 52.7–83.6%, significantly higher than single-turn attacks, with human red-teaming pushing success rates to 75.8% on average. The key finding: translation quality is the bottleneck—poor translations actually hurt attackers, suggesting that current models may resist jailbreaks in languages their training data handles less well.