African Languages Exploit LLM Safety Gaps, Research Reveals Multilingual Jailbreaking Vulnerabilities

A recent study has uncovered significant vulnerabilities in the safety mechanisms of leading Large Language Models (LLMs) when subjected to multi-turn conversational attacks using low-resource African languages. Researchers specifically investigated Afrikaans, Kiswahili, isiXhosa, and isiZulu, finding that these languages can effectively bypass the guardrails of commercial LLMs such as ChatGPT, Claude, DeepSeek, Gemini, and Grok.

The methodology involved translating existing jailbreaking prompts and evaluating LLM responses through both automated testing and human red-teaming by native speakers. While simple, single-turn translation attacks proved largely ineffective, the multi-turn conversational approach yielded high success rates. For instance, harmful response rates in English ranged from 52.7% to 83.6%, in Afrikaans from 60.0% to 78.2%, and in Kiswahili from 41.8% to 70.9% across the tested models.

Crucially, the study highlighted the role of human red-teaming, which consistently increased jailbreak rates compared to automated methods. Across all evaluated languages, the average jailbreak rate surged from 59.8% to 75.8%. The research also identified translation quality as a critical factor, noting that poor translation limited the success of jailbreak attempts, indicating that improvements in translation accuracy could further amplify these vulnerabilities.

These findings carry significant implications for the development and deployment of AI in Africa. They demonstrate that LLM safety mechanisms are not robust across diverse linguistic contexts, potentially exposing African users to harmful content or misuse. The research underscores an urgent need for AI developers to enhance multilingual safety features, particularly for low-resource languages, to ensure equitable and secure AI experiences across the continent and prevent the weaponization of linguistic diversity against AI safety.