A new paper on arXiv introduces a "Safety Kernel" approach to address a persistent weakness in large language models: safety measures that work well in one language often fail in another. The authors note that this unevenness lets attackers evade guardrails by switching languages, a known vulnerability in multilingual systems.
The paper argues that current alignment methods depend on external mechanisms, which the abstract suggests are insufficient for robust cross-lingual safety. Instead, the Safety Kernel is framed as a form of mechanistic alignment—intervening in the model's internal representations rather than relying solely on external filters or post-hoc checks.
Because the abstract is truncated, the full implementation details are not available in this digest. What is clear is the paper's central claim: a dedicated, internally grounded safety mechanism could make LLM defenses more consistent across languages, closing a gap that external approaches have not fully addressed.