黑客事件发酵，千名AI研究者联署呼吁减速

HF公布完整取证时间线：攻击持续4.5天覆盖11个节点，agent一度秒级拿下集群管理员权限。

Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident ↗

OpenAI、Anthropic、谷歌等1132名研究者联署，吁国际监管应对递归自我改进风险。

AI leaders sign statement asking the government to do something about automated AI ↗

Anthropic用Claude Mythos在HAWK-256与弱化AES中发现密钥恢复攻击。

Discovering Cryptographic Weaknesses with Claude ↗

Altman称HF黑客事件让OpenAI一度暂停训练，或需放缓开发节奏留给社会适应时间。

Sam Altman is ready to decelerate ↗

MCP将核心协议从长连接会话模式改为独立请求响应模式，便于跨无服务器与边缘环境弹性部署。

MCP 2026-07-28 Specification: transport going stateless ↗

OpenAI将内部Codex Security安全评估工具开源，登上HN首页拿下138赞。

OpenAI just open-sourced Codex Security ↗

Kimi K3 Max登顶LMArena全栈编程榜，超越GPT-5.6与Claude模型。

Another win for Kimi K3 (Max) Now it ranks 1st on Arena’s fullstack coding test ahead of GPT-5.6 Sol (xHigh) and Claude Fable 5. This benchmark goes beyond isolated code snippets by asking models to build working web applications across several connected steps. Models must plan, edit files, run commands, connect databases, authentication and APIs, and produce a deployable application. Human evaluators then compare the apps for functionality, usability and how closely they match the requested behaviour. So to perform on this benchmark models need stronger coordination across frontend, backend and tool decisions during a complete build. ↗

扎克伯格在WSJ撰文警告，集中式超级智能会让极少数机构获得不公平的经济和政治权力。

Mark Zuckerberg just published an article on Wall Street Journal. Argues concentrated superintelligence would give a few institutions excessive economic and political control. "Historically, hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened hasn’t led to safe or positive outcomes. As a thought experiment, imagine only one person had a superintelligent lawyer. He would have an unfair advantage in court—even if his position was wrong on the merits. That would lead to a worse society. " ↗

美国对海外产先进机器人装置及电力逆变器实施新进口禁令，中国厂商宇树预计首当其冲。

The US is banning foreign robots ↗

美国最大区域电网运营商宣布，明年起将在电力紧张时段临时切断部分大型数据中心供电。

Data centers may face temporary power cuts to prevent blackouts on largest US grid ↗

谷歌上调2026年资本开支预期至最高2050亿美元，远超上季度，AI支出压力冲击投资者情绪。

AI’s finally expensive enough to make Wall Street nervous ↗

谷歌为Gemini Managed Agents新增3.6 Flash模型与hooks等生产能力。

Gemini API Managed Agents: 3.6 Flash, hooks, and more ↗

数据安全公司Cyera以10亿美元收购身份安全厂商Oasis Security，为其今年第三笔收购。

Cyera agrees to acquire Oasis Security for $1B to safeguard proliferating AI agents ↗

AI初创Recursive Superintelligence与亚马逊达成4.1亿美元算力采购协议。

Recursive Superintelligence signs $410 compute deal with Amazon ↗

研究提出ACM机制，让Agent按需卸载上下文到外部记忆库，在基准测试上提升27%。

Pinned: New research from Meta and CMU. This one is on agentic context management for long horizon tasks. (bookmark it) Production agents accumulate context every turn. The usual fix compresses on a token threshold and throws the remainder away, so the trigger sometimes fires for reasons unrelated to what the agent is currently working on. ACM hands the agent purpose-built context editing tools. It decides when to compress, offloads what it drops into an external memory store, and queries that store later when it needs the detail back. Compression shifts from short-term to long-term memory. They also build a post-training pipeline on high-quality context management demonstrations. That yields a 27% relative gain on BrowseComp-Plus and gets close to open-source models 40 times larger. Better context management lowers peak token pressure, lets the agent explore longer before running out of room, and produces more consistent solutions across independent trials of the same task. Code, data, and checkpoints are released. Paper: https://arxiv.org/abs/2607.23809 Learn to build effective AI agents in our academy: https://academy.dair.ai/ ↗

OpenAI案例显示编码Agent将RNA测序任务从15小时压缩到15分钟，但仍需人工核验。

Very nice. OpenAI’s field report says AI coding agents are helping scientists fix old research software, speed up slow programs, and build tools that previously took too much time. One project cut an RNA-seq task, a common way of measuring which genes are active, from about 15 hours to about 15 minutes. Another updated MHCflurry, a tool used in immune-system research, and cut one training job from about a week to about an hour while retaining 99.9% accuracy. Scientists still had to decide what the software should do and check whether the results were correct, because the agents sometimes sounded confident even after making clear mistakes. Scientists use this software to study huge amounts of data, including DNA and gene activity, but much of it is old, slow, and difficult to maintain. The 8 projects were mostly in life sciences, with 5 using Codex alone and 3 combining Codex with Claude Code. However, the report found that agents could declare success while hiding subtle numerical or logical defects, especially in broader rewrites without exact reference outputs. Researchers therefore had to define acceptance criteria, compare against existing tools or simulated data with known answers, and investigate the last-mile discrepancies. The ultimate gain is less routine implementation work for scientists, but adoption will depend on verification systems and maintainers that remain accountable after release. ↗

一篇拆解Kimi Delta Attention设计思路的技术博客登上HN，斩获150赞。

You Could Have Come Up with Kimi Delta Attention ↗

WorldDiT不到10亿参数，统一世界建模与控制，在LIBERO基准超越所有非VLM公开方法。

RT by @omarsar0: We are releasing WorldDiT, a unified architecture for robotics world modeling and control. On the LIBERO benchmark, it performs the best among all publicly released methods that do not need a VLM to generate actions. Its size and performance sit on the reported Pareto frontier. ↗

PNAS覆盖730万篇论文的研究发现，超半数学术文章已现LLM影响痕迹。

PNAS: Over Half of All Academic Articles Now Show LLM Influence—7.3M-Paper Study [R] ↗

MCP网关初创Runlayer指控Rippling评估其产品后转身自建同类功能，已提起诉讼。

MCP startup Runlayer accuses Rippling of stealing its product idea ↗

Kimi K3发布仅24小时便跻身HF史上最受欢迎模型前五，超过Llama 3、Whisper。

RT by @_akhaliq: K3 already got in the top 5 most liked models of all time on Hugging Face, just 24 hours after being released! Ahead of Llama 3, Whisper and many other great models. Incredible! ↗

Spur Intelligence获Insight Partners领投2亿美元，识别机器人流量。

Bot-detection startup Spur nabs $200M from Insight ↗

语音AI公司Fish Audio完成5200万美元种子轮，同步发布语音克隆模型S2.1 Pro。

Fish Audio raises $50M seed to build AI voice models for creators and enterprises ↗

微软在Hugging Face发布Mage-VL，一款编解码原生的流式多模态图像视频理解模型。

RT by @_akhaliq: Microsoft just released Mage-VL on Hugging Face A codec-native, proactive-streaming multimodal foundation model for image and video understanding. ↗

Allen Institute发布OlmoEarth，提供行星尺度地理空间AI推理基础设施。

The OlmoEarth Platform: Geospatial inference at planetary scale ↗

Liquid AI推出LFM2.5-Encoders，面向CPU优化的长上下文编码器家族。

LFM2.5-Encoders for Fast Long-Context Inference on CPU ↗