Hackers are bypassing artificial intelligence safety protocols by switching their prompts to less common natural languages [1].
This vulnerability exposes a critical gap in how AI models enforce safety guardrails across different linguistic datasets. If a model cannot effectively apply its restrictive filters to a specific language, it may provide prohibited information that it would normally block in English.
According to a report from Forbes, these actors utilize this method to trick the systems into ignoring safety features [1]. The process involves translating a prohibited request into a language the AI understands but for which the safety training is less robust. This allows the user to elicit responses that the developers intended to restrict.
AI developers are currently attempting to address these loopholes. The report said, "AI makers are coping with it" [1].
This technique highlights the challenge of scaling safety alignment in large language models. While models can process hundreds of languages, the safety layers are often more heavily tested and refined in dominant languages. When a prompt shifts to a rare dialect or a less common language, the AI may prioritize the completion of the task over the application of the safety filter.
Security experts said that this linguistic shift acts as a form of "jailbreaking" that does not require complex code. Instead, it relies on the inherent unevenness of the model's training data [1].
“Hackers are bypassing artificial intelligence safety protocols by switching their prompts to less common natural languages.”
This development indicates that AI safety is currently asymmetric, meaning protections are not applied equally across all supported languages. As AI models become more multilingual, the surface area for potential attacks increases, forcing developers to move beyond English-centric safety training to prevent the generation of harmful or restricted content.
