Simon Willison大模型深度长读1 天前

Quoting Anthropic Frontier Red Team

摘要 · We evaluate several models on 100 tasks from the [internal Binary Exploitation benchmark] (selected at random), and find that GLM-5.3 develops full control flow hijacks in 4% of the trials; Claude Mythos Preview did so in 6%. Although GLM-5.3 performs below Claude Mythos Preview here, a meaningful t

阅读原文 · Simon Willison →

大模型

We evaluate several models on 100 tasks from the [internal Binary Exploitation benchmark] (selected at random), and find that GLM-5.3 develops full control flow hijacks in 4% of the trials; Claude Mythos Preview did so in 6%. Although GLM-5.3 performs below Claude Mythos Preview here, a meaningful t

原文：https://simonwillison.net/2026/Sep/29/anthropic-frontier-red-team/

0 条评论 · 观点来自社区

评论区 0 条讨论

以 … 的身份发言⌘/Ctrl+Enter 发送

还没有评论，来抢沙发。