Low-Resource Languages Jailbreak GPT-4

Safety training has a language gap: GPT-4's safety mechanisms were primarily trained on English data, leaving the model vulnerable when harmful requests are translated into low-resource languages (languages with less digital representation and training data). This creates a systematic blind spot in AI safety.

The vulnerability is exploitable at scale: Using publicly available translation tools, anyone can bypass GPT-4's safeguards by translating unsafe prompts into low-resource languages, achieving a 79% success rate on harmful requests—matching the effectiveness of sophisticated jailbreaking attacks.

Linguistic inequality now affects everyone: While language disparities have historically disadvantaged speakers of low-resource languages, this research reveals the problem now threatens all LLM users. A weakness in one language becomes a security vulnerability for the entire system.

High-resource languages show better protection: The study found that mid- and high-resource languages (like Spanish, French, and German) had significantly lower attack success rates, demonstrating that more comprehensive safety training across languages substantially improves safeguards.

Multilingual red-teaming is essential for responsible AI: The paper calls for safety researchers to expand their testing and training efforts beyond English to include diverse languages, ensuring that AI safety measures are robust across linguistic boundaries rather than creating exploitable gaps.