The Illusion of Universal Safety

Modern LLM alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF), are predominantly trained on English-centric datasets. This research highlights that safety alignment does not generalize perfectly across linguistic boundaries. While a model may successfully refuse a harmful prompt in English, the same model often fails to recognize or block the same intent when translated into lower-resource or non-English languages. This creates a dangerous 'safety gap' where the model's safety guardrails are essentially bypassed simply by changing the input language.

Mechanisms of the Cross-Lingual Gap

The study identifies that the internal representations of safety concepts are tightly coupled with English-language training data. When a user queries the model in a different language, the model's semantic processing shifts, often moving outside the 'safe' latent space established during alignment. This phenomenon suggests that current alignment strategies are not truly learning the abstract concept of 'harm,' but rather memorizing specific English-language patterns associated with refusal. Consequently, the model remains vulnerable to adversarial attacks in non-English languages because the safety training does not effectively transfer to the model's multilingual capabilities.

Implications for AI Deployment

This research serves as a critical warning for developers building global AI applications. Relying on default safety alignment from base models is insufficient for multilingual products. Developers must implement language-specific safety testing and potentially incorporate cross-lingual safety fine-tuning to ensure consistent behavior. The findings underscore that safety is not a monolithic feature of a model but a language-dependent property that requires rigorous, multi-language evaluation to prevent exploitation in international markets.