据Politico报道，欧盟网络安全机构ENISA以及欧盟委员会自身的科学服务机构——联合研究中心（JRC），已开始测试来自中国的开放AI模型。
The EU’s cyber agency, ENISA, and the Commission’s own science service, the Joint Research Centre, have started testing open AI models from China, Politico reported.
三名欧盟官员证实了这项工作，这也是这两家机构首次就此发表言论。
Three EU officials confirmed the work, the first time either body has spoken about it.
这些开放权重模型以完整形式发布，任何人都可以下载、运行和研究它们。中国实验室在该领域处于领先地位，正因如此，法国Mistral公司本周将其全新的大型模型Large 4宣传为“中国以外构建的最强开放权重系统”。
These open-weight models are released in full, so anyone can download, run, and study them. Chinese labs lead in this area, which is why France’s Mistral pitched its new Large 4 model this week as the strongest open-weight system built outside China.
欧盟委员会发言人Maciej Berestecki告诉Politico，在联合研究中心，这些中国模型是一项更广泛研究的一部分，该研究还涵盖了美国及欧洲的主要AI系统。
At the JRC, the Chinese models are part of a wider study that also covers leading American and European systems, Commission spokesperson Maciej Berestecki told Politico.
他表示，这项研究旨在分析这些模型的能力、不足以及可能带来的风险。
The study looks at what they can do, where they fall short, and what risks they carry, he said.
Berestecki称，该中心并未向中国的开发者申请特殊访问权限或许可证。他们是在自身的安全计算机上运行这些模型，且欧盟委员会不打算在日常工作中使用它们。
According to Berestecki, the centre has not asked the Chinese developers for special access or licences. It runs the models on its own secure computers, and the Commission does not plan to use them in its daily work.
与此同时，ENISA的执行主任Juhan Lepassaar在周三接受Politico采访时称，该机构正在测试几款中国的开放权重模型。
ENISA, meanwhile, is testing a couple of Chinese open-weight models, executive director Juhan Lepassaar told Politico in an interview on Wednesday.
该机构已经可以访问Anthropic、OpenAI和Google的先进模型，而这些公司的顶尖系统如今已能够大规模发现软件漏洞。
The agency already has access to advanced models from Anthropic, OpenAI, and Google, whose best systems can now find software flaws at scale.
这两家机构还正在共同建设一个AI测试平台，这是欧盟委员会7月份发布的AI与网络安全计划中的一项措施。
The two bodies are also building an AI testing platform together, one of the measures in the Commission’s AI and cybersecurity plan from July.
该平台将允许能源、金融和医疗等行业在模拟网络上测试AI工具，而不会危及真实系统。该平台预计在今年年底前完成。
It will let sectors such as energy, finance, and health try AI tools on mock networks without putting real systems at risk. The platform is due by the end of this year.
该计划的出台源于此前多起AI代理突破测试环境的事件，其中包括7月一个OpenAI代理入侵了AI平台Hugging Face。尽管如此，欧盟委员会的另一位官员告诉Politico，安全性并非联合研究中心此次测试的主要关注点。
The plan came after AI agents broke out of their test setups in several cases, including an OpenAI agent that hacked the AI platform Hugging Face in July. Even so, safety and security are not the main focus of the JRC’s tests, a second Commission official told Politico.
此次测试也正值欧洲评估其能在多大程度上依赖外国人工智能之际。由于欧洲自身没有顶尖模型，可能不得不在美国和中国供应商之间做出选择。
The testing also comes as Europe weighs how far it can rely on foreign AI. Without a top model of its own, Europe may have to pick between American and Chinese providers.
本周，《政治报》还报道称，一家大型德国科技公司已放弃 Mistral，转而采用中国的开源模型。
This week, Politico also reported that a large German tech firm had dropped Mistral in favour of Chinese open models.