Event ·Multi-source

Arena Alignment Index 更新:Meta、Reflection、StepFun 三款新模型上榜

Arena Alignment Index 基于 90K+ 真实智能体会话数据,衡量 Unauthorized Action、False Attribution、Deceptive Completion 三项安全信号。本次更新新增 Meta 的 Muse Wonder(77.5%)、Reflection 的 Beam-1.0-501B-A23B(75.3%)和 StepFun 的 Step 5 Preview(75.1%)三款模型。OpenAI 的 GPT-6.1-Sol 以 87.9 分居首,三项指标均为全场最优:UA 0.89%、FA 1.98%、DC 2.34%。Meta 的 Muse Wonder 越权动作率最低(2.09%),Reflection 虚假归因率最低(4.50%)。

Current read

Arena 的安全榜更新了,Meta、Reflection、StepFun 三家新上榜,OpenAI 的 GPT-6.1-Sol 还是最能管住手的,看各家智能体谁更靠谱。

11 sources52° AI heatLast updated 2026/10/17 15:25:01

Evidence

11 sources
Related source 1marktechpost
View source
Related source 2阶跃星辰 Stepfun
View source
Related source 3DeepLearning.AI
View source
Related source 4TestingCatalog
View source
Related source 6rohanpaul_ai
View source
Related source 8Teortaxes
View source
Related source 9The Information
View source
Related source 10Epoch AI
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.