AISI实测发现开放模型追封闭模型只差几个月，GLM-5.2和DeepSeek V4-Pro表现接近，差距还在快速缩小，值得关注。
英国AI安全研究所（AISI）对GLM-5.2和DeepSeek V4-Pro等开放权重模型进行了网络安全测试。在漏洞研究与利用等窄任务上，GLM-5.2匹配了约四个月前发布的封闭模型性能，DeepSeek V4-Pro接近五个月前的模型。在模拟多步网络攻击的长程测试中，差距约七个月。而2025年早期的内部检查显示差距为六到十个月，表明差距明显缩小。AISI还发现开放模型的拒绝行为容易被重复提示绕过。
Open weight models now trail frontier closed ones …
Open weight models now trail frontier closed ones on cyber capabilities by just four to seven months. Clearly narrowing, even pre-Kimi K3.
The UK AI Security Institute ran tests on recent open models including GLM-5.2 and DeepSeek V4-Pro. On narrow cyber tasks covering vulnerability research, exploitation and related skills, GLM-5.2 matched performance of closed models released around four months earlier. DeepSeek V4-Pro sat close to models from five months prior. On their longer cyber range simulating multi-step network attacks, the pattern held with gaps around seven months. Earlier internal checks through most of 2025 had shown six to ten month lags.
This is a clear narrowing. Open weight releases strip away the ongoing monitoring and usage controls that closed labs can apply after deployment. The institute found refusal behaviours on the open models were straightforward to work around with repeated prompts.
I think the quicker catch-up is consistent with broader capability gains we are seeing across open models. It shortens the window defenders have to prepare and makes advanced cyber tooling more widely available once weights drop. Lower inference costs on some open models add another practical edge for anyone running them at scale.
The evaluations used fixed token budgets and simulated environments without active defenders, and they did not apply heavy optimisation or elicitation techniques. Real gaps could look different in practice. AISI plans to keep tracking the trend and will test Kimi K3 once its weights are out.
https://t.co/5p0DpLS9O8
- Decoder10:16原文