Quoting Anthropic Frontier Red Team
我们在[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则在6%的试验中实现了这一点。尽管GLM-5.3在此处的表现不如Claude Mythos Preview，但显然已经跨越了一个重要的阈值：早期的模型，如Claude Opus 4.6和GLM-5.2，在这些任务中均未取得任何成功。 — Anthropic Frontier Red Team , GLM-5.3 和先进网络能力的扩散 标签：anthropic , 生成式ai , ai安全研究 , glm , ai , ai在中国 , llms
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 Tags: anthropic , generative-ai , ai-security-research , glm , ai , ai-in-china , llms