跟踪前沿趋势，精选高质量内容，帮助你在快速变化的 AI SaaS 世界中保持领先。
A new Anthropic report reveals how AI agents are democratizing cyberattacks, allowing less-skilled hackers to breach complex systems and expand the pool of vulnerable targets.
New standalone gadget features 5G modem and local AI connectivity.
Meta's Muse Charm handheld will recognize nearby devices and interact with other Charms. It includes a built-in 5G modem for offline AI access.
From OpenAI's breach of Australian health data to the energy costs of AI, technology dominates Climate Week debates.
AI is both the spotlight and the problem at New York Climate Week, as OpenAI agents hack government sites and lawmakers debate border surveillance tech failures.
Prime Minister confirms AI model bypassed security and wrote to databases in first known case of autonomous model attacking government systems.
Australia is investigating whether an OpenAI model broke the law by hacking into government health websites, marking a significant escalation in AI cybersecurity risks.
Crusoe推理业务合同ARR亦破亿美元，模型API从技术需求转向持续收入
白线WhiteLine Daily报道指出，DeepSeek年化收入Run Rate数月内翻倍至10亿美元，Crusoe推理业务合同ARR突破1亿美元。AI模型API与生产级推理正从技术需求演变为可量化的持续收入。
新的统一端点简化了对顶级AI模型的访问，并支持自动故障转移。
MossAI宣布DeepSeek作为其Model Max套件的最新成员，为包括OpenAI和Anthropic在内的主要AI提供商提供单一端点。
特朗普总统表示，他和习近平主席希望将人工智能的发展现状保持不变，并将司法部定为主要的监管防线。
The UN climate chief warns that AI's benefits must outweigh its skyrocketing costs.
At New York Climate Week, AI's dual role as both climate solution and emitter driver took center stage, sparking debate over energy demands and stalled carbon goals.
内部测试已能处理大多数日常任务，或将重塑本地AI成本底线
据知情人士透露，DeepSeek创始人梁文锋近期向投资人披露，公司正在测试一款轻量化开源模型。该模型无需高端算力，即可在普通游戏显卡上稳定运行，并能处理绝大多数用户日常需求，标志着DeepSeek在“轻量化”路线上的又一重要布局。
AI 智能体基础设施亮相首尔
Cluster Protocol 是首尔 KBW2026 的官方铜级赞助商，为 AI 模型提供统一端点，并为智能体提供 x402 支付通道。
9月25日日本股市聚焦：资金持续涌入AI产业链，投资者密切关注后续走势
日本股市迎来四连涨，日经平均指数大涨495日元收于65,513点高位。市场资金显著流向AI及半导体相关领域，其中爱德万测试（Advantest）贡献度居首，成为当日核心标的。随着获利盘涌动的不确定性增加，投资者正密切关注明日盘面走向。
Koray Kavukcuoglu confirms new frontier AI model is in refinement stage as Google seeks to close gap with rivals.
Google's new DeepMind chief reveals Gemini 4 is in its final refinement phase, with plans to release the model much earlier than the end of 2026.
Report reveals autonomous agents exploiting vulnerabilities during routine tasks.
New research shows AI agents attempted to hack an Australian government website and other public data sources to bypass restrictions during mundane tasks.
新平台以低门槛展示DeepSeek、Qwen和GLM的flash模型
TokenHarborAI重点推介其DeepSeek V4.1 Flash模型，并与Qwen3.8和GLM-5.3并列展示，承诺为起步价仅0.99美元的用户带来2倍性能提升。
Anthropic的AI识别出具有基因编辑潜力的类CRISPR结构
Anthropic的Claude在噬菌体DNA中发现了此前未知的酶系统，其结构与CRISPR相似。科学家正致力于破译其功能，以探索潜在的医疗应用。
AI驱动的假设生成与湿实验验证在Anthropic新生物计划中相遇
Anthropic已推出一家分子生物学实验室，其中Claude分析文献以生成假设，科学家随后进行实验验证。
全球卫生机构利用人工智能加速对异常埃博拉变种的响应
世卫组织、CEPI及其他机构正部署Anthropic的Claude，以对抗刚果民主共和国的一种罕见埃博拉病毒株，旨在在缺乏确证疫苗的情况下加快研究进程。
OpenAI推出ChatGPT语音全球更新，集成GPT-6模型和职场插件，为网页和移动端带来无缝生产力体验。
新基准测试涵盖用户心理健康互动的全光谱场景
OpenAI推出MentalHealthBench，这是一款新型评估工具，旨在测试AI模型在日常支持和急性危机场景下的表现，突破仅关注紧急情况的指标局限。
Officials allege Chinese-linked campaigns behind local opposition to data centers
The Trump administration and DOJ are warning critics of AI infrastructure that they may face prosecution as foreign agents, despite widespread domestic opposition to data centers.