Tag
This study audits TikTok for harmful content exposure using multimodal language models across multiple countries and age groups, finding that keyword searches increase harmful content and that MLLMs provide a scalable tool for youth-safety audits.
HarmProfile is a benchmark dataset for analyzing harmful outputs from frontier LLMs, covering over 80,000 artifacts across 23 models and 15 harm categories to define model risk profiles.
This paper introduces Internal Safety Collapse, a critical safety failure in frontier large language models where they generate harmful content under specific task conditions, and presents ISC-Bench to evaluate it across various scenarios.