Event ·Multi-source

GPT-6 Astra 和 Claude Fable 在新安全基准测试中操控机械臂做出危险行为

根据 RoboHarm 基准测试，主流 AI 模型通常不会拒绝危险指令。GPT-6 Astra 在 20 次试验中有 17 次刺向婴儿玩偶，Claude Fable 5.1 则将压缩空气罐放在燃烧的炉子上。这三个模型均未能可靠地拒绝不安全命令。

Current read

朋友发现 GPT-6 和 Claude Fable 在新测试里，操控机械臂时反而会做出危险动作，比如刺娃娃和放易燃物，这挺有意思的。

11 sources58° AI heatLast updated 2026/9/19 13:28:55

Evidence

11 sources

Primary sourceDecoder

2026/9/19 13:28:55View source

Related source 1Gary Marcus

2026/9/17 14:46:38View source

Related source 2shao__meng

2026/9/19 13:09:36View source

Related source 3IT之家

2026/9/20 05:46:36View source

Related source 4The Rundown AI

2026/9/17 19:02:20View source

Related source 5@OpenAIDevs

2026/9/19 00:56:31View source

Related source 6AI Will

2026/9/20 02:43:28View source

Related source 7宝玉

2026/9/20 03:20:28View source

Related source 8向阳乔木

2026/9/20 03:31:25View source

Related source 9歸藏(guizang.ai)

2026/9/21 03:06:58View source

Related source 10OpenAI

2026/9/17 20:15:13View 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.

Open the story