Anthropic Frontier Red Team 对 GLM-5.3 和高级网络能力的研究

Quoting Anthropic Frontier Red Team

AI 导读

Anthropic Frontier Red Team 对 GLM-5.3 和 Claude Mythos Preview 在内部二进制利用基准测试中的表现进行了评估，发现 GLM-5.3 在 4% 的测试中实现了完整的控制流劫持，而 Claude Mythos Preview 在 6% 的测试中实现了。

原文

29th September 2026

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 threshold has clearly been crossed: earlier models, like Claude Opus 4.6 and GLM-5.2, do not succeed in any of them.

— Anthropic Frontier Red Team, GLM-5.3 and the spread of advanced cyber capabilities

来源：Simon Willison · simonwillison.net