Erfan Shayegani - Visual Grounding Safety in Vision-Language Models

​Can an AI refuse a harmful request in words, yet comply by pointing? Erfan Shayegani presents “Visual Grounding Safety in Vision-Language Models,” developed during his recent Apple internship. The talk explores safety gaps when models locate objects in images, and how targeted training can improve safety while preserving their visual capabilities.

Erfan Shayegani is a fourth-year Computer Science PhD student at UC Riverside researching multimodal AI safety and trustworthy agents. He has held research internships at Apple and Microsoft Research.