Anthropic 红队：GLM-5.3 与高级网络能力扩散
原标题：Quoting Anthropic Frontier Red Team
AI 摘要
Anthropic Frontier Red Team 在内部 Binary Exploitation 基准的 100 个随机任务上评估多个模型，发现 GLM-5.3 在 4% 的试验中实现了完整的控制流劫持，Claude Mythos Preview 为 6%。报告指出，尽管 GLM-5.3 表现低于 Claude Mythos Preview，但已明显跨过一个有意义的门槛：此前的 Claude Opus 4.6 和 GLM-5.2 在这些任务中均未成功。
正文节选
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 advance