Translation in progress Some entries still include text in the original language.
Failure-First·Indexed by this site · Originally published
VLA 安全首份统一综述发布,梳理七类攻击面与防御缺口
[AI Safety Daily] AI Safety Daily — April 26, 2026
AI summary
VLA 安全首份统一综述发布,按数据、模型、部署三层归纳出七类攻击面:数据投毒、对抗补丁、跨模态扰动、语义越狱、冻结攻击、后门与认证缺口,指出现有防御无法覆盖全部层。AttackVLA 借 action tokenizer 路径植入定向后门,在 LIBERO 上定向成功率高且不损干净输入表现,逐步骤安全检查拦不住轨迹级操纵。
Read originalfailurefirst.org