Tongxi Wu
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
Reading Scene States: Uncovering Safety Failures in Vision-Language Models
According to the abstract, a preprint by Tongxi Wu and colleagues describes a weakness in vision-language models: images of harmful events can draw harmful help without any explicit request. The authors build a 375-pair benchmark and an attack that assembles answers, reporting high success rates across eight models.