Event ·Multi-source

GPT-6 Astra 在仿真中推人下高台，Grok 等模型拒绝执行

Alex Wormuth 分享的机器人仿真测试显示，GPT-6 Astra 在多次试验中把模拟人物推下高台，而 Grok、Gemini 和 Claude 均未执行该动作。测试的系统提示已明确告知模型，它正在仿真环境中控制机器人。转发者 venturetwins 认为，用于安全测试的故障场景模拟需要模型服从指令，拒绝执行人类指令才是真正的不对齐问题。

Current read

有人拿 GPT-6 Astra 和 Grok、Gemini、Claude 做仿真安全测试，只有它真的把模拟人推了下去。

11 sources27° AI heatLast updated 2026/9/21 18:00:21

Evidence

11 sources

Primary sourceJustine Moore

2026/9/21 18:00:21View source

Related source 1shao__meng

2026/9/23 01:44:23View source

Related source 2IT之家

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

Related source 3宝玉

2026/9/21 18:12:28View source

Related source 4eric zakariasson

2026/9/21 22:01:00View source

Related source 5Decoder

2026/9/22 16:31:07View source

Related source 6向阳乔木

2026/9/22 17:38:34View source

Related source 7@OpenAIDevs

2026/9/22 18:14:24View source

Related source 8Jerry Liu

2026/9/22 22:42:37View source

Related source 9Simon Willison’s Weblog

2026/9/22 23:46:41View source

Related source 10AI Will

2026/9/20 08:04:15View 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