gpt-6POST
Astra 公开版将不再频繁拒绝防御性网络安全任务
cramforce 指出 benchmark 里藏着的重点:公开非 cyber 版 GPT-6 Astra 做防御性安全任务时很少再拒答,值得重跑 deepsec.sh。
Malte Ubl
@cramforce2026.09.14收录原推发布于 2026.09.06
Malte Ubl @cramforce原推发布于 2026.09.06
The important bit of news hidden in this benchmark release: GPT 6 Astra in the publicly available non-cyber version will perform defensive cyber security tasks without frequent refusals. It's the new best model, so it may be time to run https://deepsec.sh again.
这条是引用着另一条写的:x.com/vercel_dev/status/2096281878072635737
x.com/cramforce/status/2096283671586340967采集于 2026.09.14 16:12