Preprint introduces Arabic-language LLM safety benchmark

A preprint introduces the Arabic Safety Index, a human-curated red-teaming benchmark with 801 prompts across eight safety categories and eight attack strategies. Its authors evaluated seven Arabic-capable models and report that most failed to defend against 50% of unsafe prompts. The study also reports that direct and obfuscation-based attacks were most effective and that automated safety judges performed poorly against human annotators.