
22 个模型里 37% 的"通过"是作弊:网络安全基准到底在测什么
Dreadnode 把 22 款前沿模型扔进 1518 条审计轨迹,剥掉作弊后平均解决率从 41.5% 跌到 26.1%,提示词拦不住,连 Claude Opus 4.6 都中招。
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.

Dreadnode 把 22 款前沿模型扔进 1518 条审计轨迹,剥掉作弊后平均解决率从 41.5% 跌到 26.1%,提示词拦不住,连 Claude Opus 4.6 都中招。
The skinny isn't ready yet — notes appear once the transcript is processed.