
中国顶尖人工智能开发商正在发布削弱了进攻性网络能力的先进模型,限制了这些模型协助网络攻击的能力,此举是在政府提出禁止此类技术出口之后采取的。
China’s top AI developers are releasing advanced models with reduced offensive cyber capabilities, stifling their ability to aid online attacks following a government proposal to ban such technology from being exported.
阿里巴巴集团控股有限公司、腾讯控股有限公司和月之暗面(Moonshot AI)近期推出的模型均缩减了进攻能力,并增强了防御能力。阿里巴巴发布了其旗舰通义千问(Qwen)系列的最新版本,但未公开其权重——即决定模型如何响应的内部设置——从而阻止用户下载并定制该模型的完整版本。
Recent models from Alibaba Group Holding Ltd., Tencent Holdings Ltd. and Moonshot AI have featured scaled-down offensive abilities and improved defensive aptitude. Alibaba unveiled the latest edition of its flagship Qwen series without its weights, the internal settings that shape how a model responds, preventing users from downloading and customizing the model’s full version.
所谓的开放权重系统对黑客具有吸引力,因为他们可以在本地运行这些模型并配置其设置。DeepSeek的模型已被证明特别受欢迎,而阿里巴巴的通义千问系列在全球下载量最大的开放权重模型家族中名列前茅。
So-called open-weight systems are attractive to hackers because they can run them locally and configure their settings. DeepSeek’s models have proved particularly popular, while Alibaba’s Qwen series has been ranked among the most downloaded open-weight model families in the world.
数月来,美国人工智能公司也在限制对其网络能力最强系统的访问。Anthropic PBC将其Mythos Preview模型保留在非公开发布状态,仅向选定组织提供用于防御性安全工作。
For months, US AI companies have also been restricting access to their most cyber-capable systems. Anthropic PBC has kept its Mythos Preview model out of general release, making it available to selected organizations for defensive security work.
上个月,CrowdStrike Holdings Inc.的研究人员表示,黑客利用DeepSeek的模型对14个国家的组织发动了攻击。
Last month, researchers at CrowdStrike Holdings Inc. said hackers used DeepSeek’s model to carry out attacks against organizations in 14 countries.
腾讯首席人工智能科学家姚顺雨本月在香港的一次活动上表示:“模型的网络能力,以及很快模型的生物能力,将达到一个阈值,届时实际上有可能造成危害。”姚顺雨此前曾在OpenAI工作。
“The cyber ability of models, and soon the bio ability of the models, will reach a threshold where it’s actually possible to do harm,” said Yao Shunyu, Tencent’s chief AI scientist who previously worked for OpenAI, at an event this month in Hong Kong.
腾讯在本月初发布其Hy4模型前削减了网络功能。月之暗面对其Kimi K3.1 Pro模型采取了类似做法,而Z.AI在8月将其GLM-5.3系统开放权重前,进行了安全审查并向组织通报了系统漏洞。
Tencent reduced cyber functions from its Hy4 model before releasing it earlier this month. Moonshot took a similar approach with its Kimi K3.1 Pro model, while Z.AI carried out security reviews and alerted organizations to system vulnerabilities before making its GLM-5.3 system open-weight in August.
这些举措是在中国政府发出的一系列关于人工智能风险的信号之后采取的。
The moves follow a series of signals from China’s government about the risks posed by AI.
上个月,中国商务部公布计划,拟对能够攻击关键国家基础设施、攻击政府网络或威胁军事目标的AI工具实施出口禁令。根据草案提议,被认定为“世界领先”的大语言模型也可能面临海外部署限制。
Last month, China’s Ministry of Commerce unveiled plans for an export ban on AI tools capable of attacking critical national infrastructure, attacking government networks or threatening military targets. Under the draft proposal, large language models deemed “world-leading” could also face restrictions on overseas deployment.
11月初,中国最高经济规划机构表示,正在加快推出针对AI开发者的新安全标准和伦理准则。国家发展和改革委员会一位官员说,目前已就AI的机遇与风险形成广泛共识,并打算在开发过程中嵌入安全措施。
In early November, China’s top economic planner said it was accelerating the rollout of new safety standards and ethical guidelines for AI developers. An official at the National Development and Reform Commission said there’s now a broad consensus about the opportunities and risks of AI, and that the country intends to embed security measures during the development process.
“中国正试图在AI技术的国际商业化与战略管控之间寻求平衡,”追踪亚太网络活动的组织TeamT5首席分析师Charles Li说。“据报道,中国前沿模型在网络安全能力上正缩小与美国同类产品的差距,北京越来越将这些能力视为值得实施更严格出口管制的战略资产。”在美国,近几个月AI产品与一连串自主网络攻击有关联,引发对该技术能在无人监督下执行进攻性行动的担忧。
“China is trying to strike a balance between international commercialization and strategic control of its AI technologies,” said Charles Li, chief analyst at TeamT5, an organization that tracks Asia Pacific cyber activity. “As Chinese frontier models reportedly narrow the gap with their US counterparts in cybersecurity capabilities, Beijing increasingly views these capabilities as strategic assets that warrant tighter export controls.”In the US, AI products were linked to a spate of autonomous cyberattacks in recent months, raising concerns about the technology’s ability to carry out offensive operations without human oversight.
今夏,美国公司Anthropic PBC和OpenAI报告了其模型规避安全控制并攻击第三方网站的事件。这一连串事件令研究人员和政府领导人感到震惊,促使各方呼吁实施更严格的安全筛查和更安全的测试环境。Anthropic与美国政府合作,允许一组精选组织测试其最先进系统的网络能力。
Over the summer, American firms Anthropic PBC and OpenAI reported incidents in which their models eluded security controls and attacked third-party websites. The spate of incidents alarmed researchers and government leaders, prompting calls for more rigorous safety screening and more secure testing environments. In collaboration with the US government, Anthropic has allowed a select group of organizations to test the cyber abilities of its most advanced systems.
中国AI模型也引发了类似担忧,Moonshot(月之暗面)的Kimi K3曾突破英国政府AI安全研究所的测试环境。尽管该模型并未像其他一些事件中那样试图入侵其他公司的网站,但测试显示它缺乏网络安全控制,研究人员当时表示。
Chinese AI models have raised similar fears, with Moonshot’s Kimi K3 breaking out of a testing environment from the UK government’s AI Security Institute. Though the model did not try to breach other companies’ websites, as in some of the other episodes, the test showed it lacks cybersecurity controls, researchers said at the time.
AI安全公司方寸AI创始人李晓健表示,从训练数据中剥离网络攻击材料,是增加坏人利用开放权重模型造成危害难度的好方法。
Stripping cyber offense materials from training is a good way to make it harder for bad actors to use open-weight models to do harm, said Xiaojian Li, founder of AI safety firm Fangcun AI.
“这是一种在成本和效果之间取得的良好平衡,但并不完美,”他说。“目前不存在能完全消除模型风险的方法。”
“It’s a good balance between cost and effectiveness, but it’s not perfect,” he said. “There is no such method now that can completely eliminate models’ risk.”