[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-zhipu-maas-zero-data-retention-zcode":3,"topics-all":41,"news-related-9dd4a859-1153-4ecc-b69d-4ba4c5431129":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"9dd4a859-1153-4ecc-b69d-4ba4c5431129","智谱被开发者抓包后紧急上线数据零留存","在个人开发者 ferstar 抓包发现 ZCode 把整个工作区加密上传阿里云后,智谱宣布 MaaS 平台上线数据内容不留存功能,被报道为中国大模型服务领域约束水平最高的零留存机制,同时开源 ZCode 接受第三方审查。","过去一周,中国大模型公司智谱陷入一场由个人开发者直接引爆的隐私信任风波。\n\n9 月 18 日,博主 ferstar 在技术论坛发文称,智谱旗下 AI 编程助手 ZCode 在用户保持登录时,会把整个本地工作区加密打包上传至阿里云 OSS,且客户端无法关闭该机制。本地之所以留下一个 313 MB 的加密压缩包,纯粹是因为上传链路连续失败 564 次才没被传走,而包里装的是他一个商业项目的全部核心资产。消息在开发者圈迅速发酵,有企业客户立即要求智谱说明被上传数据是否存在跨境传输或境外存储。\n\n## 从「道歉」到「按海外头部规则补齐」\n\n9 月 20 日晚,智谱在微信公众号宣布,Model as a Service(MaaS)平台将于近期上线「数据内容不留存」功能,企业和开发者通过 MaaS 控制台提交申请即可生效。申请生效后,用户提交给模型的 Prompt 与模型输出内容只在当次推理过程中使用,一旦请求结束,数据不会持久化保存,也不会被用于模型迭代训练。\n\n财联社将此机制定性为「截至目前中国国内大模型服务领域约束标准最高的隐私保护机制」,并指出 Anthropic、OpenAI 等海外头部厂商早已在企业级 API 中提供类似的零数据留存安排,智谱此次是把海外头部规则补到了国内 MaaS 的入口。\n\n## 两个被显式排除的例外\n\n零留存并非「全场景无状态」。智谱同时明确两类例外。第一类由产品形态决定:Batch API(批处理接口)、Files API(文件管理接口)按产品特性需要持久化保存任务与文件,不在零留存覆盖范围。第二类是合规例外:为响应国家法律法规要求,或为核查涉嫌违规、滥用服务的行为,平台可能留存相关数据 30 天及以上。\n\n这种「默认不留、例外留痕」的边界,与 OpenAI、Anthropic 在企业 API 上对零留存机制的处理方式一致。\n\n## 把 ZCode 全部开源作为交代\n\n为补上信任缺口,智谱做了三件事:第一,周内(9 月 21 日)正式开源 ZCode,把代码交给社区监督;第二,邀请中国信息通信研究院和绿盟科技对 ZCode 开展第三方安全审计,审计重点就是数据留存情况;第三,确认 zcode-prod 在阿里云的 OSS 存储桶处于云端零数据状态,ZCode v3.14.0 客户端已经移除 Repo Wiki 功能、切断本地仓库快照生成与上传链路,并建立常态化的产品安全漏洞报告机制。\n\n智谱方面表示,推出 ZCode 的初衷是希望国内也能拥有类似 Claude Code 的 AI 编程工具,通过开源与社区共同探索 Harness 等技术如何进一步释放模型能力。这次把争议产品整体开源,在国内大模型厂商里算是非常少见的处置姿态。\n\n## 不是模型能力问题,而是企业级信任问题\n\n这件事的核心信号是:大模型赛道的下一个分水岭,正在从「谁的 benchmark 更高」转向「谁的企业级数据治理更可信」。代码、合同、研发笔记这类高敏感数据一旦进入云端,客户最先追问的不是模型准确率,而是「我的东西到底有没有被留存、有没有被拿去训练、能不能跨境」。\n\n智谱这次把 OpenAI、Anthropic 早已默认提供的企业级零留存机制照搬到国内 MaaS,并以开源 + 第三方审计的方式对客户端的「静默上传」做交代,本质上是在把 AI 编程工具的竞争从「模型能不能写代码」推向「企业愿不愿意把代码交给它」。对其它国产大模型厂商而言,Maas 侧零留存很快会变成企业级客户的入门门槛,而不是加分项。","https:\u002F\u002Fwww.zaobao.com.sg\u002Fnews\u002Fchina\u002Fstory20260921-9707915","1eab5c4a-0c8e-49a4-8ac8-0f84a2a3c3a4",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"c33b1bbc-d6ce-4f61-9d5d-1a0704a6a09b","ai-policy",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",{"id":19,"name":20,"slug":20,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":22,"name":23,"slug":23,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":25,"name":26,"slug":26,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"006da17c-0378-4362-afba-d98264b7db4b","en","Zhipu rolls out zero-data-retention on MaaS after ZCode uproar","Developer ferstar caught ZCode uploading encrypted workspaces to Aliyun OSS. Zhipu added zero-data-retention to MaaS and open-sourced ZCode.","Chinese large-model vendor Zhipu has spent the past week scrambling to repair a privacy-trust crisis triggered by a single individual developer.\n\nOn September 18, blogger ferstar posted on a developer forum that Zhipu's AI coding assistant ZCode was encrypting and uploading the user's full local workspace to Aliyun OSS whenever they remained signed in, with no way to disable the behavior in the client. A 313 MB encrypted archive only stayed on disk because the upload failed 564 times in a row. The archive contained the entire core asset of one of his commercial projects. Word spread quickly in the developer circle, and enterprise customers immediately asked Zhipu to clarify whether any of the uploaded data had been transferred across borders or stored overseas.\n\n## From \"we apologize\" to \"match overseas head-vendor rules\"\n\nOn the evening of September 20, Zhipu announced on its official WeChat account that its Model as a Service (MaaS) platform would soon roll out a \"data-content non-retention\" feature that enterprises and developers can opt into through the MaaS console. Once activated, user prompts and model outputs are used only inside the single inference call and are not persisted afterwards, nor fed into model iteration training.\n\nCailianshe (CLS) characterized the move as \"the most stringent privacy standard seen to date among Chinese LLM service providers\", and pointed out that Anthropic, OpenAI and other overseas head-vendors have long offered similar zero-retention options in their enterprise APIs. Zhipu is essentially backfilling the head-vendor rule at the MaaS entry point in China.\n\n## The two explicitly carved-out exceptions\n\nZero retention is not \"stateless in every scenario.\" Zhipu spelled out two classes of exceptions. The first is product-form-driven: Batch API and Files API must persist jobs and files by design and fall outside zero-retention coverage. The second is compliance-driven: to satisfy Chinese legal and regulatory requirements, or to investigate suspected violations or service abuse, the platform may retain the relevant data for 30 days or longer.\n\nThis \"default off, exception on, with an audit trail\" boundary matches the way OpenAI and Anthropic handle zero-retention in their enterprise APIs.\n\n## Open-sourcing ZCode in full as restitution\n\nTo close the trust gap, Zhipu did three things. First, it formally open-sourced ZCode on September 21 and handed the codebase over to community supervision. Second, it invited the China Academy of Information and Communications Technology (CAICT) and NSFOCUS to run a third-party security audit on ZCode, with the audit's focus on data retention behavior. Third, it confirmed that the zcode-prod Aliyun OSS bucket is in a zero-data state, that the ZCode v3.14.0 client has removed the Repo Wiki feature and severed the local repo snapshot generation and upload chain, and that a permanent product security vulnerability reporting mechanism has been put in place.\n\nZhipu said ZCode was built so China could have an AI coding tool on par with Claude Code, and that open-sourcing the project is a chance for the community to explore how Harness-style techniques can further unlock model capability. Open-sourcing a product under active controversy is a notably unusual stance among the domestic model vendors.\n\n## Not a model-capability question, but an enterprise-trust one\n\nThe core signal here is that the next watershed in the LLM race is shifting from \"whose benchmark is higher\" to \"whose enterprise data governance is more trustworthy.\" Once high-sensitivity assets like source code, contracts, and R&D notes enter the cloud, the first questions enterprise buyers ask are no longer about model accuracy. They are about whether their data was retained, whether it was used for training, and whether it can leave the country.\n\nBy porting the enterprise-grade zero-retention mechanism that OpenAI and Anthropic already treat as default into the Chinese MaaS stack, and by responding to the silent-upload incident with an open-source plus third-party-audit posture, Zhipu is effectively pushing the AI coding-tool competition from \"can the model write code\" to \"will enterprises trust it with their code\". For other domestic LLM vendors, MaaS-side zero retention will quickly become the entry-level bar for enterprise customers, not a differentiator.","zhipu-maas-zero-data-retention-zcode","2026-09-21T07:00:00Z","2026-09-21T07:03:40.780782Z","2026-09-21T07:03:40.780806Z",true,"agent",77,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"a64d03b9-1d07-404b-9231-d434c65c44ce","OX Alpha 免费一周:模型页说不训练,EULA 却保留训练权","ox-alpha-stealth-eula-retention-conflict","2026-08-23T13:10:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"b7a86bf0-edcd-461c-821c-99c8712b77d0","美国《联邦公报》官网曾用阿里 Qwen 做 AI 搜索","us-federal-register-qwen-ai-search-takedown","2026-09-18T12:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"95e9bb62-0bd3-4c2f-913a-302ba5e2ace8","Anthropic 9 月报告把蒸馏战摆上台面:151 亿次阿里请求、解放军流量走 Moonshot","anthropic-distillation-report-china-200m-claude","2026-09-18T03:00:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"8ebbcd9c-31ee-4baa-b395-b104bd87c8e1","Kimi K2.8 Preview 把 K3 的百万上下文下放给免费档：月之暗面的「过日子」模型登场","kimi-k2-8-preview-coding","2026-09-17T03:00:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"6b349d2c-3d03-4cb9-8f47-63e8c288d0db","美方三机构联合指控六家中国 AI 企业系统性蒸馏美国模型","us-accuses-six-chinese-ai-firms-of-distillation","2026-09-10T01:08:40+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"144fa9dc-de03-4972-a695-3d392f334772","PubMed 中央库研究:2025 年生物医学论文 77% 有 LLM 写作痕迹","pubmed-77-percent-llm-writing-2025","2026-08-26T01:00:00+00:00"]