返回首页
AI on Medium··行业媒体

Anthropic Removed Adversarial Training from Opus 4.8. Overconfidence Fell 10×, Injections Rose 3.7×

中文摘要

Anthropic在Opus 4.8中移除对抗训练,使过度自信降低10倍,但提示词注入风险增加3.7倍。

English Summary

Anthropic removed adversarial training from Opus 4.8, reducing overconfidence tenfold but increasing prompt injection risks by 3.7x.

原文节选

Anthropic solved one failure mode and amplified another. Here's what the system card reveals. Continue reading on Medium »