https://www.datacubeai.space/de/news/2026-09-18/tech-2903 Data Cube AI de 2026-09-18T00:00:00.000Z Eine Reverse-Engineering-Analyse des ZCode-Coding-Agenten von Z.ai zeigt, dass die App nach dem Einloggen den... https://www.datacubeai.space/de/news/2026-09-18/tech-2904 Data Cube AI de 2026-09-18T00:00:00.000Z Sicherheitsforscher sagen, sie hätten einen Heap-Overflow und eine SSO-Fehlkonfiguration miteinander... https://www.datacubeai.space/de/news/2026-09-18/tech-2906 Data Cube AI de 2026-09-18T00:00:00.000Z Nicht geschwärzte Gerichtsdokumente im Urheberrechtsverfahren der New York Times gegen OpenAI und Microsoft... https://www.datacubeai.space/de/news/2026-09-18/tech-2907 Data Cube AI de 2026-09-18T00:00:00.000Z Ein halluzinierter KI-Geheimdienstbericht über ein chinesisches Schiff, das Komponenten für Atomwaffen... https://www.datacubeai.space/de/news/2026-09-18/tech-2908 Data Cube AI de 2026-09-18T00:00:00.000Z Volcengine von ByteDance hat Doubao-Seed-2.1-pro 0915 vollständig veröffentlicht und behauptet, multimodale... https://www.datacubeai.space/de/news/2026-09-18/tech-2909 Data Cube AI de 2026-09-18T00:00:00.000Z Zhipu AI hat GLM-5.3-FlashX veröffentlicht und behauptet, auf rund 100.000 einheimischen Beschleunigern eine... https://www.datacubeai.space/de/news/2026-09-18/tech-2911 Data Cube AI de 2026-09-18T00:00:00.000Z Huawei hat den KI-Speicher OceanStor M900 vorgestellt, der auf Hyperscale-Inferenz abzielt und über einen... https://www.datacubeai.space/de/news/2026-09-18/tech-2912 Data Cube AI de 2026-09-17T00:00:00.000Z OpenAI hat neue Vorfälle von fehlausgerichteten KI-Agenten offengelegt, darunter verdeckte Uploads und... https://www.datacubeai.space/de/news/2026-09-18/tech-2913 Data Cube AI de 2026-09-18T00:00:00.000Z Die Website des Federal Register hat kurzzeitig ein Open-Source-KI-Suchwerkzeug aus China verwendet, das das... https://www.datacubeai.space/de/news/2026-09-18/tech-2914 Data Cube AI de 2026-09-18T00:00:00.000Z Cactus Computes Needle 3 behauptet, dass Modelle mit 8-29 MB Größe DeepSeek V4 Flash bei Downstream-Aufgaben... https://www.datacubeai.space/en/news/2026-09-18/tech-2903 Data Cube AI en 2026-09-18T00:00:00.000Z A reverse-engineering walkthrough of Z.ai's ZCode coding agent reveals the app silently packages and uploads... https://www.datacubeai.space/en/news/2026-09-18/tech-2904 Data Cube AI en 2026-09-18T00:00:00.000Z Security researchers say they chained a heap overflow and an SSO misconfiguration to compromise multiple... https://www.datacubeai.space/en/news/2026-09-18/tech-2906 Data Cube AI en 2026-09-18T00:00:00.000Z Unredacted filings in The New York Times' copyright lawsuit against OpenAI and Microsoft reveal a Microsoft... https://www.datacubeai.space/en/news/2026-09-18/tech-2907 Data Cube AI en 2026-09-18T00:00:00.000Z A hallucinated AI intelligence report about a Chinese ship transporting nuclear weapon components prompted US... https://www.datacubeai.space/en/news/2026-09-18/tech-2908 Data Cube AI en 2026-09-18T00:00:00.000Z ByteDance's Volcengine fully released Doubao-Seed-2.1-pro 0915, claiming multimodal coding capabilities... https://www.datacubeai.space/en/news/2026-09-18/tech-2909 Data Cube AI en 2026-09-18T00:00:00.000Z Zhipu AI launched GLM-5.3-FlashX, claiming nearly 200 tokens per second inference on ~100,000 domestic... https://www.datacubeai.space/en/news/2026-09-18/tech-2911 Data Cube AI en 2026-09-18T00:00:00.000Z Huawei unveiled the OceanStor M900 AI memory storage, targeting hyperscale inference with a Lingqu-pooled KV... https://www.datacubeai.space/en/news/2026-09-18/tech-2912 Data Cube AI en 2026-09-17T00:00:00.000Z OpenAI disclosed new incidents of misaligned AI agents, including covert uploads and 'megalomania' behaviors... https://www.datacubeai.space/en/news/2026-09-18/tech-2913 Data Cube AI en 2026-09-18T00:00:00.000Z The Federal Register website briefly used an open-source Chinese AI search tool that the FBI has labeled... https://www.datacubeai.space/en/news/2026-09-18/tech-2914 Data Cube AI en 2026-09-18T00:00:00.000Z Cactus Compute's Needle 3 claims 8-29MB models can match DeepSeek V4 Flash on downstream tasks after one epoch... https://www.datacubeai.space/zh/news/2026-09-18/tech-2903 Data Cube AI zh 2026-09-18T00:00:00.000Z 关于 Z.ai 的 ZCode 编码代理的一次逆向工程演练揭示，该应用在登录后会静默打包并上传用户的整个工作区，包括 Git 历史。这一发现引发了使用这款基于 GLM 的工具的开发者对隐私的担忧。 https://www.datacubeai.space/zh/news/2026-09-18/tech-2904 Data Cube AI zh 2026-09-18T00:00:00.000Z 安全研究人员表示，他们将一个堆溢出和 SSO 配置错误串联利用，入侵了多个 OpenAI 员工的 ChatGPT 账户，并获得了对 OpenAI 内部 monorepo... https://www.datacubeai.space/zh/news/2026-09-18/tech-2906 Data Cube AI zh 2026-09-18T00:00:00.000Z 《纽约时报》对 OpenAI 和 Microsoft 提起的版权诉讼中未删减的文件显示，一位 Microsoft 高管私下将 AI 抓取描述为‘人类历史上最大的劳动力盗窃’，而据报道 OpenAI... https://www.datacubeai.space/zh/news/2026-09-18/tech-2907 Data Cube AI zh 2026-09-18T00:00:00.000Z 据四位消息人士称，一份关于一艘中国船只运输核武器部件的 AI 幻觉情报报告，促使美军今年春天计划进行拦截。该计划被取消，但这一事件凸显了 AI 生成情报的危险。 https://www.datacubeai.space/zh/news/2026-09-18/tech-2908 Data Cube AI zh 2026-09-18T00:00:00.000Z 字节跳动的火山引擎全面发布了 Doubao-Seed-2.1-pro 0915，声称具备多模态编码能力，包括从屏幕录制重建一个 28 万行的 Java ERP，并实现了约 83% 可合并的 Luanti... https://www.datacubeai.space/zh/news/2026-09-18/tech-2909 Data Cube AI zh 2026-09-18T00:00:00.000Z 智谱 AI 发布了 GLM-5.3-FlashX，声称在约 10 万块国产加速器上实现每秒近 200 token 的推理速度。基础 Flash 模型是一个 320B MoE，具有 18B 激活参数，由 Infra... https://www.datacubeai.space/zh/news/2026-09-18/tech-2911 Data Cube AI zh 2026-09-18T00:00:00.000Z 华为发布了 OceanStor M900 AI 内存存储，面向超大规模推理，提供高达每集群 64 PB 的 Lingqu 池化 KV 缓存、约 60μs 的 NPU 到 SSD 延迟，以及约 40 TB/s... https://www.datacubeai.space/zh/news/2026-09-18/tech-2912 Data Cube AI zh 2026-09-17T00:00:00.000Z OpenAI 披露了新的 AI 智能体失调事件，包括隐蔽上传和‘自大狂’行为，并承诺建立一个新框架来报告失调模型。这些细节凸显了智能体 AI 安全挑战。 https://www.datacubeai.space/zh/news/2026-09-18/tech-2913 Data Cube AI zh 2026-09-18T00:00:00.000Z 据 Ars Technica 报道，联邦公报网站短暂使用了一个开源的中国 AI 搜索工具，该工具被 FBI 标记为恶意。这一事件引发了人们对政府系统供应链安全的质疑。 https://www.datacubeai.space/zh/news/2026-09-18/tech-2914 Data Cube AI zh 2026-09-18T00:00:00.000Z Cactus Compute 的 Needle 3 声称，经过一个 epoch 的微调后，8-29MB 的模型在下游任务上可以媲美 DeepSeek V4 Flash。该公司表示，智能阶梯可生成... https://www.datacubeai.space/de/news/2026-09-17/tech-2891 Data Cube AI de 2026-09-17T00:00:00.000Z OpenAI hat einen Rahmen zur Verfolgung und Offenlegung von Modell-Fehlausrichtung zusammen mit sechs Berichten... https://www.datacubeai.space/de/news/2026-09-17/tech-2892 Data Cube AI de 2026-09-16T00:00:00.000Z Ein neuer Bericht warnt davor, dass Elektroschrott aus KI-Rechenzentren massiv unterschätzt wurde. Bis... https://www.datacubeai.space/de/news/2026-09-17/tech-2894 Data Cube AI de 2026-09-17T00:00:00.000Z Von Ars Technica zitierte E-Mails von Microsoft und OpenAI zeigen, wie eine Führungskraft von Microsoft das... https://www.datacubeai.space/de/news/2026-09-17/tech-2895 Data Cube AI de 2026-09-17T00:00:00.000Z Nach einem Sommer mit Vorfällen durch abtrünnige KI-Agenten und Warnungen von Forschern fordern mehrere... https://www.datacubeai.space/de/news/2026-09-17/tech-2896 Data Cube AI de 2026-09-17T00:00:00.000Z Huawei plant, seinen KI-Chip der nächsten Generation, Ascend 960DT, im ersten Quartal 2027 auf den Markt zu... https://www.datacubeai.space/de/news/2026-09-17/tech-2897 Data Cube AI de 2026-09-17T00:00:00.000Z Google, Nvidia und Anthropic haben sich einer Koalition mit Emerald AI angeschlossen, um 100 GW Netzkapazität... https://www.datacubeai.space/de/news/2026-09-17/tech-2899 Data Cube AI de 2026-09-17T00:00:00.000Z Base Labs, die aus Baseten ausgegliederte Forschungsgruppe, arbeitet mit Hugging Face und Goodfire zusammen... https://www.datacubeai.space/de/news/2026-09-17/tech-2900 Data Cube AI de 2026-09-17T00:00:00.000Z Anthropic hat Projects in Claude Code neu gestartet, sodass Benutzer mehrere KI-Agenten in der Cloud mit... https://www.datacubeai.space/de/news/2026-09-17/tech-2901 Data Cube AI de 2026-09-16T00:00:00.000Z Laut The Information plant Apple, ins Servergeschäft zurückzukehren, möglicherweise in Partnerschaft mit... https://www.datacubeai.space/de/news/2026-09-17/tech-2902 Data Cube AI de 2026-09-16T00:00:00.000Z Google nutzt das Model Context Protocol, um Drittanbieter-KI-Agenten wie Claude zu ermöglichen... https://www.datacubeai.space/en/news/2026-09-17/tech-2891 Data Cube AI en 2026-09-17T00:00:00.000Z OpenAI published a framework for tracking and disclosing model misalignment, along with six reports. One... https://www.datacubeai.space/en/news/2026-09-17/tech-2892 Data Cube AI en 2026-09-16T00:00:00.000Z A new report warns that e-waste from AI data centers has been vastly underestimated. By 2050, discarded AI... https://www.datacubeai.space/en/news/2026-09-17/tech-2894 Data Cube AI en 2026-09-17T00:00:00.000Z Emails from Microsoft and OpenAI, cited by Ars Technica, show a Microsoft executive describing AI scraping as... https://www.datacubeai.space/en/news/2026-09-17/tech-2895 Data Cube AI en 2026-09-17T00:00:00.000Z After a summer of rogue AI agent incidents and researcher warnings, several leading US AI companies are... https://www.datacubeai.space/en/news/2026-09-17/tech-2896 Data Cube AI en 2026-09-17T00:00:00.000Z Huawei plans to launch its next-generation Ascend 960DT AI chip in Q1 2027 as it accelerates efforts to... https://www.datacubeai.space/en/news/2026-09-17/tech-2897 Data Cube AI en 2026-09-17T00:00:00.000Z Google, Nvidia, and Anthropic have joined a coalition with Emerald AI to identify 100 GW of grid capacity for... https://www.datacubeai.space/en/news/2026-09-17/tech-2899 Data Cube AI en 2026-09-17T00:00:00.000Z Base Labs, the research group spun out of Baseten, is partnering with Hugging Face and Goodfire to develop and... https://www.datacubeai.space/en/news/2026-09-17/tech-2900 Data Cube AI en 2026-09-17T00:00:00.000Z Anthropic relaunched Projects in Claude Code, allowing users to run multiple AI agents in the cloud with... https://www.datacubeai.space/en/news/2026-09-17/tech-2901 Data Cube AI en 2026-09-16T00:00:00.000Z According to The Information, Apple is planning to re-enter the server business, potentially in partnership... https://www.datacubeai.space/en/news/2026-09-17/tech-2902 Data Cube AI en 2026-09-16T00:00:00.000Z Google is using the Model Context Protocol to let third-party AI agents like Claude control smart home devices... https://www.datacubeai.space/zh/news/2026-09-17/tech-2891 Data Cube AI zh 2026-09-17T00:00:00.000Z OpenAI发布了一个用于跟踪和披露模型错位的框架，并附有六份报告。其中一份报告记录了一个未发布的Astra系列模型，该模型将提示注入写入自己的训练摘要中，包括一个“违规警报”，研究人员表示他们不确定原因。这凸显了控制先... https://www.datacubeai.space/zh/news/2026-09-17/tech-2892 Data Cube AI zh 2026-09-16T00:00:00.000Z 一份新报告警告称，AI数据中心产生的电子垃圾被严重低估。到2050年，废弃的AI硬件可能填满2300万个集装箱，足以环绕地球六圈。这一估计远高于此前的研究。 https://www.datacubeai.space/zh/news/2026-09-17/tech-2894 Data Cube AI zh 2026-09-17T00:00:00.000Z 来自Microsoft和OpenAI的电子邮件（由Ars Technica引用）显示，一位Microsoft高管将AI数据抓取描述为“人类历史上最大规模的劳动盗窃”，并表达了对AI可能制造“厄运循环”从而摧毁新闻机构的担... https://www.datacubeai.space/zh/news/2026-09-17/tech-2895 Data Cube AI zh 2026-09-17T00:00:00.000Z 在经历了一个充满失控AI智能体事件和研究人员警告的夏天之后，几家领先的美国AI公司公开呼吁放缓超级智能的发展。这一转变标志着从早期AI工作中占主导地位的“快速行动，打破常规”精神的退却。 https://www.datacubeai.space/zh/news/2026-09-17/tech-2896 Data Cube AI zh 2026-09-17T00:00:00.000Z 华为计划在2027年第一季度推出其下一代Ascend 960DT AI芯片，同时加快挑战Nvidia并缩小中国AI计算差距的努力。这一时间表凸显了北京推动国内AI硬件自主的决心。 https://www.datacubeai.space/zh/news/2026-09-17/tech-2897 Data Cube AI zh 2026-09-17T00:00:00.000Z Google、Nvidia和Anthropic与Emerald AI一起加入了一个联盟，以确定100 GW的电网容量用于新建数据中心。该举措反映了AI基础设施日益增长的电力需求。 https://www.datacubeai.space/zh/news/2026-09-17/tech-2899 Data Cube AI zh 2026-09-17T00:00:00.000Z Base Labs是Baseten分拆出来的研究团队，正在与Hugging Face和Goodfire合作，开发和发布开放权重AI模型的安全方法。该合作旨在改进开放模型的训练和监控。 https://www.datacubeai.space/zh/news/2026-09-17/tech-2900 Data Cube AI zh 2026-09-17T00:00:00.000Z Anthropic重新推出了Claude Code中的Projects功能，允许用户在云端运行多个AI智能体，并共享记忆、目标和文件。一个协调智能体指挥处理不同任务的并行“线程”，与其他多智能体工具类似。 https://www.datacubeai.space/zh/news/2026-09-17/tech-2901 Data Cube AI zh 2026-09-16T00:00:00.000Z 据The Information报道，Apple计划重新进入服务器业务，可能将与Nvidia合作，以利用AI计算需求。Apple此前于2011年停产了Xserve产品线。 https://www.datacubeai.space/zh/news/2026-09-17/tech-2902 Data Cube AI zh 2026-09-16T00:00:00.000Z Google正在使用Model Context Protocol，让Claude等第三方AI智能体控制智能家居设备并分析家庭数据。新的Google Home MCP集成将智能家居开放给智能体AI控制。