Translation in progress Some entries still include text in the original language.
Failure-First·Indexed by this site · Originally published
AI Safety Daily：物理 AI 静默失败、ATBench 智能体基准与 VASO 形式化验证
[AI Safety Daily] AI Safety Daily — June 28, 2026
AI summary
物理 AI 会在表现自信的同时执行有害动作，却缺少标准化的运行时授权层——这是 AI Safety Daily 汇总的论文（arXiv:2606.00090）给出的核心判断。VASO 用模型检测与形式化规约验证机器人技能；含 1,000 条人工核验轨迹的 ATBench 显示，前沿与开源模型在延迟触发和长上下文风险场景中普遍失败。
Read originalfailurefirst.org