Anthropic红队报告：GLM-5.3与Claude
Quoting Anthropic Frontier Red Team
前沿模型安全能力的量化突破，直接反映AI在二进制漏洞利用等高危场景的实际威胁水平，值得安全与模型研究者重点关注。
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.
我们在来自[内部二进制利用基准测试]（随机选取）的100个任务上评估了多个模型，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在此项指标上的成功率为6%。尽管GLM-5.3在此处的表现低于Claude Mythos Preview，但一个有意义的阈值显然已被跨越：早期的模型，如Claude Opus 4.6和GLM-5.2，在这些任务中均未取得成功。
— Anthropic Frontier Red Team, GLM-5.3 and the spread of advanced cyber capabilities
—— Anthropic前沿红队，GLM-5.3与高级网络能力的扩散
Tags: anthropic, generative-ai, ai-security-research, glm, ai, ai-in-china, llms
标签：anthropic, generative-ai, ai-security-research, glm, ai, ai-in-china, llms
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力