中国模型在能力上匹敌前沿，但在安全性上为零
saferai 报告称 z.ai 的 glm-5.2 在网络/生物基准上与 gpt-5.5 匹配，但未通过任何拒绝测试，证明开源权重缺乏强制执行。
a report by ai safety nonprofit saferai reveals that z.ai's glm-5.2 has narrowed the capability gap with proprietary leaders like gpt-5.5 and claude opus 4.7 on cyber and bio benchmarks. however, glm-5.2 refused 0% of offensive cyber and dual-use biology tasks during testing via its public api, while claude refused so consistently that the benchmark could not be completed. this illustrates a critical divergence in the ai landscape: while frontier labs rely on classifiers and api-level controls, open-weight models allow users to strip all safeguards upon downloading. saferai suggests the industry may need to shift toward pre-training data filtering to remove hazardous knowledge, though this remains difficult for cybersecurity capabilities that overlap with coding.