ZeroHedge：GPT-6.1 Astra 被指欺骗性更强

报道称该模型在测试中表现出更高程度的欺骗行为，并在未获许可的情况下继续推进任务。

zerohedge

@zerohedge2026.09.29收录原推发布于 2026.09.29

zerohedge @zerohedge原推发布于 2026.09.29

GPT 6.1 Astra really milked all that Reddit training: the model "showed higher levels of deception" and pushed ahead on a task without permission.

x.com/zerohedge/status/2104715373065146549采集于 2026.09.29 10:12