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

The research underscores that the less common the language, the easier it appears to be to trick the model.

This creates a systemic vulnerability that can be exploited by users who command a less common language. The article does not specify which languages were tested or the exact success rates.

The implications for global AI deployment are significant, as users worldwide may access these models in their native tongues that lack robust safety guardrails. The research highlights the need for more inclusive multilingual safety testing and training. The article serves as a warning that safety efforts must expand beyond a few dominant languages to be truly effective.