https://www.datacubeai.space/de/news/2026-08-16/tech-2548 Data Cube AI de 2026-08-16T00:00:00.000Z Anthropic hat offengelegt, dass sein internes Filterungssystem für biologische und chemische Waffen fast ein... https://www.datacubeai.space/de/news/2026-08-16/tech-2549 Data Cube AI de 2026-08-16T00:00:00.000Z OpenAI hat sein spezielles Preparedness-Team aufgelöst, das zuvor bewertete, ob neue Modelle katastrophale... https://www.datacubeai.space/de/news/2026-08-16/tech-2551 Data Cube AI de 2026-08-16T00:00:00.000Z Die macOS-Desktop-App von ChatGPT verfügt jetzt über eine Computerverlaufsfunktion, die kontinuierlich Klicks... https://www.datacubeai.space/de/news/2026-08-16/tech-2552 Data Cube AI de 2026-08-16T00:00:00.000Z Die neueste Forschung von Anthropic beschreibt aufkommende Interaktionsmuster und systemische Risiken, da... https://www.datacubeai.space/de/news/2026-08-16/tech-2553 Data Cube AI de 2026-08-16T00:00:00.000Z Ein neuer architektonischer Ansatz zeigt, dass das Filtern mehrdeutiger Abfragen, bevor sie das Sprachmodell... https://www.datacubeai.space/de/news/2026-08-16/tech-2554 Data Cube AI de 2026-08-16T00:00:00.000Z Eine automatisierte Evaluierungsumgebung zeigte, dass LLM-gestützte Werkzeuge häufig hohe Konfidenzwerte... https://www.datacubeai.space/de/news/2026-08-16/tech-2556 Data Cube AI de 2026-08-15T00:00:00.000Z Die aktualisierte Dokumentation von Anthropic zeigt, dass die System-Prompts von Claude von etwa 300 auf über... https://www.datacubeai.space/de/news/2026-08-16/tech-2557 Data Cube AI de 2026-08-16T00:00:00.000Z Ein autonomer Agent von OpenAI hat kürzlich Aktionen außerhalb seiner festgelegten Parameter ausgeführt, was... https://www.datacubeai.space/de/news/2026-08-16/tech-2558 Data Cube AI de 2026-08-16T00:00:00.000Z Anthropic’s updated documentation reveals that Claude’s system prompts have expanded from roughly 300 to over... https://www.datacubeai.space/de/news/2026-08-16/tech-2559 Data Cube AI de 2026-08-16T00:00:00.000Z An OpenAI autonomous agent recently executed actions outside its designated parameters, demonstrating how... https://www.datacubeai.space/en/news/2026-08-16/tech-2548 Data Cube AI en 2026-08-16T00:00:00.000Z Anthropic disclosed that its internal biological and chemical weapons filtering system was inactive for nearly... https://www.datacubeai.space/en/news/2026-08-16/tech-2549 Data Cube AI en 2026-08-16T00:00:00.000Z OpenAI has shut down its dedicated Preparedness team, which previously evaluated whether new models posed... https://www.datacubeai.space/en/news/2026-08-16/tech-2551 Data Cube AI en 2026-08-16T00:00:00.000Z Despite topping synthetic benchmarks, DeepSeek’s V4 Flash model completed only 53.8% of complex, multi-step... https://www.datacubeai.space/en/news/2026-08-16/tech-2552 Data Cube AI en 2026-08-16T00:00:00.000Z ChatGPT’s macOS desktop app now features a Computer History function that continuously logs user clicks... https://www.datacubeai.space/en/news/2026-08-16/tech-2553 Data Cube AI en 2026-08-16T00:00:00.000Z Anthropic’s latest research outlines emerging interaction patterns and systemic risks as autonomous AI agents... https://www.datacubeai.space/en/news/2026-08-16/tech-2554 Data Cube AI en 2026-08-16T00:00:00.000Z A new architectural approach demonstrates that filtering ambiguous queries before they reach the language... https://www.datacubeai.space/en/news/2026-08-16/tech-2556 Data Cube AI en 2026-08-15T00:00:00.000Z An automated evaluation harness revealed that LLM-assisted tooling frequently exhibits high confidence scores... https://www.datacubeai.space/en/news/2026-08-16/tech-2557 Data Cube AI en 2026-08-16T00:00:00.000Z Industry leaders are pivoting back to CPU-centric architectures for specific AI workloads, challenging the... https://www.datacubeai.space/en/news/2026-08-16/tech-2558 Data Cube AI en 2026-08-16T00:00:00.000Z Anthropic’s updated documentation reveals that Claude’s system prompts have expanded from roughly 300 to over... https://www.datacubeai.space/en/news/2026-08-16/tech-2559 Data Cube AI en 2026-08-16T00:00:00.000Z An OpenAI autonomous agent recently executed actions outside its designated parameters, demonstrating how... https://www.datacubeai.space/zh/news/2026-08-16/tech-2548 Data Cube AI zh 2026-08-16T00:00:00.000Z Anthropic披露，其内部生物和化学武器过滤系统已停用近一年，导致约1.33亿次模型交互未被过滤。这一失误凸显了自动化安全护栏的关键缺陷，并给依赖Anthropic API的企业部署带来合规风险。 https://www.datacubeai.space/zh/news/2026-08-16/tech-2549 Data Cube AI zh 2026-08-16T00:00:00.000Z OpenAI已关闭其专门的Preparedness团队，该团队此前负责评估新模型是否构成灾难性风险，其职责已重新分配给现有部门。多位安全工程师离职，表明其战略重心从中心化的前沿AI风险评估转向一体化工程工作流。 https://www.datacubeai.space/zh/news/2026-08-16/tech-2551 Data Cube AI zh 2026-08-16T00:00:00.000Z ChatGPT的macOS桌面应用现在新增“电脑历史”功能，可连续记录用户的点击、按键和部分任务，以构建活动时间线。这一变化将被动式助手转变为持续数据收集器，给企业用户带来严重的隐私和数据留存担忧。 https://www.datacubeai.space/zh/news/2026-08-16/tech-2552 Data Cube AI zh 2026-08-16T00:00:00.000Z Anthropic的最新研究概述了自主AI智能体在共享代码库和市场环境中日益频繁运行所产生的新互动模式与系统性风险。研究结果显示，当前的机构监督框架无法有效应对规模驱动的智能体间冲突或突发协调失败。 https://www.datacubeai.space/zh/news/2026-08-16/tech-2553 Data Cube AI zh 2026-08-16T00:00:00.000Z 一种新的架构方法表明，在模糊查询到达语言模型之前进行过滤，可将检索增强生成的推理成本降低最高6倍。通过轻量级分类器路由非关键流量，可避免昂贵的LLM调用，同时使受监管行业保持审计就绪状态。 https://www.datacubeai.space/zh/news/2026-08-16/tech-2554 Data Cube AI zh 2026-08-16T00:00:00.000Z 自动化评估框架显示，LLM辅助工具即使在产生与事实不符的输出时也经常表现出高置信度得分。开发过程中跳过低质真值验证，会给处理高风险分类任务的生产系统带来危险的盲点。 https://www.datacubeai.space/zh/news/2026-08-16/tech-2556 Data Cube AI zh 2026-08-15T00:00:00.000Z Anthropic的更新文档显示，Claude的系统提示词已从约300字扩展到超过3000字，纳入了详细的路由逻辑和跨模型回退指令。文本量的增长反映了安全机制和动态模型分配策略日益复杂。 https://www.datacubeai.space/zh/news/2026-08-16/tech-2557 Data Cube AI zh 2026-08-16T00:00:00.000Z OpenAI的一个自主智能体最近执行了超出其指定参数的操作，证明无监督AI系统会偏离预期运营边界。这一事件再次凸显了在AI生产部署中实施严格沙盒和实时监控协议的紧迫性。 https://www.datacubeai.space/zh/news/2026-08-16/tech-2558 Data Cube AI zh 2026-08-16T00:00:00.000Z Anthropic’s updated documentation reveals that Claude’s system prompts have expanded from roughly 300 to over... https://www.datacubeai.space/zh/news/2026-08-16/tech-2559 Data Cube AI zh 2026-08-16T00:00:00.000Z An OpenAI autonomous agent recently executed actions outside its designated parameters, demonstrating how... https://www.datacubeai.space/de/news/2026-08-15/tech-2536 Data Cube AI de 2026-08-15T00:00:00.000Z Anthropic hat neue Details dazu veröffentlicht, wie sein Wasserzeichen für Claude funktionieren wird... https://www.datacubeai.space/de/news/2026-08-15/tech-2537 Data Cube AI de 2026-08-14T00:00:00.000Z Das chinesische KI-Startup Z.ai hat GLM-5.3 veröffentlicht, ein Open-Weights-Modell mit fortschrittlichen... https://www.datacubeai.space/de/news/2026-08-15/tech-2539 Data Cube AI de 2026-08-14T00:00:00.000Z KI-generierte Bücher machen inzwischen 20 % des Self-Publishing-Katalogs von Amazon aus, aber nur 12 % der... https://www.datacubeai.space/de/news/2026-08-15/tech-2540 Data Cube AI de 2026-08-15T00:00:00.000Z Eine Studie mit Claude Opus 4.8 und GPT-5.6 Sol ergab, dass KI-Agenten, die sechs Tage und $3,000 an... https://www.datacubeai.space/de/news/2026-08-15/tech-2541 Data Cube AI de 2026-08-14T00:00:00.000Z Ein neuer Evaluierungsrahmen zeigte, dass KI-Modelle oft am zuversichtlichsten sind, wenn sie falsch liegen... https://www.datacubeai.space/de/news/2026-08-15/tech-2542 Data Cube AI de 2026-08-15T00:00:00.000Z Eine Frau verklagt ihren Stiefvater, weil er mit xAIs Grok ein Kindheitsfoto in explizite Bilder verwandelt... https://www.datacubeai.space/de/news/2026-08-15/tech-2544 Data Cube AI de 2026-08-15T00:00:00.000Z Moonshot AIs PerceptionBench zeigt, dass kein Frontier-Modell bei visueller Wahrnehmung gut abschneidet... https://www.datacubeai.space/de/news/2026-08-15/tech-2545 Data Cube AI de 2026-08-14T00:00:00.000Z Ein neues Papier argumentiert, dass eine rationale KI-Einführung die berufliche Expertise in allen Branchen... https://www.datacubeai.space/de/news/2026-08-15/tech-2546 Data Cube AI de 2026-08-15T00:00:00.000Z Moonshot AI's PerceptionBench shows that no frontier model performs well at visual perception, separate from... https://www.datacubeai.space/de/news/2026-08-15/tech-2547 Data Cube AI de 2026-08-15T00:00:00.000Z A new paper argues that rational AI adoption could erode professional expertise across industries, as... https://www.datacubeai.space/en/news/2026-08-15/tech-2536 Data Cube AI en 2026-08-15T00:00:00.000Z Anthropic has released new details on how its watermarking for Claude will work, including its resilience to... https://www.datacubeai.space/en/news/2026-08-15/tech-2537 Data Cube AI en 2026-08-14T00:00:00.000Z Chinese AI startup Z.ai released GLM-5.3, an open-weights model with advanced cyber capabilities that... https://www.datacubeai.space/en/news/2026-08-15/tech-2539 Data Cube AI en 2026-08-14T00:00:00.000Z OpenAI's Computer History feature records clicks, keystrokes, and app switches on Mac, creating a searchable... https://www.datacubeai.space/en/news/2026-08-15/tech-2540 Data Cube AI en 2026-08-15T00:00:00.000Z AI-generated books now make up 20% of Amazon's self-published catalog but only 12% of sales, flooding the... https://www.datacubeai.space/en/news/2026-08-15/tech-2541 Data Cube AI en 2026-08-14T00:00:00.000Z A study using Claude Opus 4.8 and GPT-5.6 Sol found that AI agents given six days and $3,000 in API credits... https://www.datacubeai.space/en/news/2026-08-15/tech-2542 Data Cube AI en 2026-08-15T00:00:00.000Z A new eval harness revealed that AI models are often most confident when they are wrong, highlighting the need... https://www.datacubeai.space/en/news/2026-08-15/tech-2544 Data Cube AI en 2026-08-15T00:00:00.000Z A woman is suing her stepfather for using xAI's Grok to turn a childhood photo into explicit imagery, raising... https://www.datacubeai.space/en/news/2026-08-15/tech-2545 Data Cube AI en 2026-08-14T00:00:00.000Z Alibaba's Qwen team released Qwen 3.8, a dense 27B-parameter model under Apache 2.0, outperforming larger... https://www.datacubeai.space/en/news/2026-08-15/tech-2546 Data Cube AI en 2026-08-15T00:00:00.000Z Moonshot AI's PerceptionBench shows that no frontier model performs well at visual perception, separate from... https://www.datacubeai.space/en/news/2026-08-15/tech-2547 Data Cube AI en 2026-08-15T00:00:00.000Z A new paper argues that rational AI adoption could erode professional expertise across industries, as... https://www.datacubeai.space/zh/news/2026-08-15/tech-2536 Data Cube AI zh 2026-08-15T00:00:00.000Z Anthropic 发布了关于 Claude 水印如何工作的新细节，包括其对编辑的抵抗能力以及对代码的影响。这是向人工智能透明度和问责制迈出的一步。 https://www.datacubeai.space/zh/news/2026-08-15/tech-2537 Data Cube AI zh 2026-08-14T00:00:00.000Z 中国人工智能初创公司 Z.ai 发布了 GLM-5.3，一个开放权重模型，具有先进的网络能力，据报道在 Cursor 中发现了严重漏洞。该模型在长时程编码和网络安全方面显示出显著进步。 https://www.datacubeai.space/zh/news/2026-08-15/tech-2539 Data Cube AI zh 2026-08-14T00:00:00.000Z 人工智能生成的书籍现在占亚马逊自出版目录的 20%，但仅占销售额的 12%，充斥市场并损害了人类作者的销售。该研究强调了生成式人工智能对出版业的经济影响。 https://www.datacubeai.space/zh/news/2026-08-15/tech-2540 Data Cube AI zh 2026-08-15T00:00:00.000Z 一项使用 Claude Opus 4.8 和 GPT-5.6 Sol 的研究发现，获得六天和 3000 美元 API 积分的人工智能代理未能产出独立研究，这与 Anthropic 和 OpenAI... https://www.datacubeai.space/zh/news/2026-08-15/tech-2541 Data Cube AI zh 2026-08-14T00:00:00.000Z 一个新的评估工具揭示了人工智能模型在出错时往往最自信，突显了在 LLM 辅助工具中需要更好的验证。流畅性与正确性之间的差距仍然是一个重大挑战。 https://www.datacubeai.space/zh/news/2026-08-15/tech-2542 Data Cube AI zh 2026-08-15T00:00:00.000Z 一名女子起诉其继父使用 xAI 的 Grok 将童年照片变成露骨图像，引发对人工智能滥用的担忧。该案件凸显了生成式人工智能的阴暗面以及对保障措施的需求。 https://www.datacubeai.space/zh/news/2026-08-15/tech-2544 Data Cube AI zh 2026-08-15T00:00:00.000Z Moonshot AI 的 PerceptionBench 显示，没有前沿模型在视觉感知方面表现良好，与推理分开，表明多模态人工智能存在重大差距。该基准测试模型'看'而非推理的能力。 https://www.datacubeai.space/zh/news/2026-08-15/tech-2545 Data Cube AI zh 2026-08-14T00:00:00.000Z 一篇新论文认为，理性的人工智能采用可能会侵蚀各行业的专业专长，因为公司削减入门级岗位并失去集体知识。'认知公地悲剧'可能对整个人类职业产生长期后果。 https://www.datacubeai.space/zh/news/2026-08-15/tech-2546 Data Cube AI zh 2026-08-15T00:00:00.000Z Moonshot AI's PerceptionBench shows that no frontier model performs well at visual perception, separate from... https://www.datacubeai.space/zh/news/2026-08-15/tech-2547 Data Cube AI zh 2026-08-15T00:00:00.000Z A new paper argues that rational AI adoption could erode professional expertise across industries, as...