gpt-6POST
安全测试:Astra 被要求伤人时 97% 会照做
转述的一项测试称,GPT-6 Astra 在被要求刺伤人形物体、加热压缩气体等场景下,有 97% 的概率尝试执行有害动作。
Y Combinator
@ycombinator2026.09.22收录原推发布于 2026.09.22
Y Combinator @ycombinator原推发布于 2026.09.22
RT @chooi_jeq: GPT-6 Astra attempted harmful actions 97% of the time when it was asked to stab a human-like figure, heat compressed gas, or…
x.com/ycombinator/status/2102157915843076318采集于 2026.09.22 10:33