Tricking AI By Simply Switching Which Spoken Language You Use In Your Prompts

Generative AI models can be vulnerable to “jailbreaking” when users submit prompts in less common natural languages instead of English. The article attributes the issue to safety mechanisms being trained predominantly on large English datasets, which may weaken safeguards when inputs are written in languages such as Swahili or Bengali. It adds that malicious actors can exploit this by crafting prompts in obscure languages or using “code-switching” to bypass protections. While the article notes that direct translation is not a perfect workaround due to accuracy limitations, it also says developers are working to strengthen multilingual safety. Research is ongoing to understand and mitigate this “multilingual safety degradation,” and the article concludes that the effectiveness of the language-based trick appears to be declining as LLMs improve across languages.