Fine-tuning attacks in one language can break multilingual safety alignment, suggesting safety information is partly language-agnostic.
Oct 1, 2024
