Report 2026-06-13 146 sources Input Tokens: 88,591 · Output Tokens: 3,151 · Total Tokens: 93,012

AI 戰情報告 (2026-06-13)

anthropicgooglenvidiaspacexhuaweiawsopenaimoonshot ai claude-fable-5mythos-5gemma-4diffusiongemmakimi-k2.7-codedeepseek-v4-pro agentsovereignty-aion-device-aiwebmcpcybersecurityipo

AI 戰情報告 (2026-06-13)

昨日 AI 重點新聞

美國政府發布出口管制，Anthropic 宣佈全球停用 Fable 5 與 Mythos 5 ref-76, ref-0, ref-99

Anthropic 在產品發佈僅 3 天 後，被迫宣佈全面停用 Claude Fable 5 與 Mythos 5。美國商務部引用國家安全法規發布出口管制指令，禁止所有外國國民（包括 Anthropic 內部非美籍研究員）存取這兩款模型。據傳是有競爭對手向政府舉報該模型存在可被利用於開發大規模網路武器的「窄域越獄 (Narrow Jailbreak)」風險。Anthropic 官方表示這是一個巨大的「誤解」，並指出 OpenAI GPT-5.5 也具備類似能力卻未受限。

Google 發表 WebMCP 標準，Chrome 149 開放 AI 代理原生調動 ref-11, ref-23

Google 提出 WebMCP (Web Model Context Protocol) 開放標準，並於 Chrome 149 啟動原型測試。該標準允許網頁開發者將特定功能直接暴露給 AI 代理，使代理能穩定地執行網頁操作（如自動填表、調用 JS 函數），而非依賴不可靠的 DOM 抓取。Google 同時宣佈 Gemini 3.5 系列將成為首批全面支援該標準的代理原生模型。

SpaceX 史上最大規模 IPO 首日大漲，Elon Musk 成為首位兆元富翁 ref-132, ref-80, ref-45

SpaceX 以 2.1 兆美元 的市值完成在 Nasdaq 的首日交易，股價收於 161.11 美元。這場歷史性 IPO 募集了 750 億美元，並讓 Elon Musk 的身價突破 1 兆美元。SpaceX 計畫利用這筆資金在全球建立由 Starlink 驅動的「太空 AI 數據中心」，目標在 2030 年前產生 3220 億美元 的 AI 相關營收。

NVIDIA Blackwell 橫掃 AgentPerf 基準測試，效能提升 20 倍 ref-24

NVIDIA 揭露 Blackwell GB300 NVL72 在業界首個代理效能基準測試 AgentPerf 中，運行 DeepSeek V4 Pro 等複雜代理任務時，每兆瓦電力能支持的代理數量比 Hopper 架構高出 20 倍。這顯示 AI 算力戰場已從單次推論轉向「長時鏈結代理流」的能效比競爭。

華為發表 HarmonyOS 7，內建 2000 個專屬 AI 代理搶攻 Apple 中國缺口 ref-52

華為發表 HarmonyOS 7，採用「代理友好 (Agent-friendly)」架構，標配升級版的語音助手與超過 2,000 個 針對特定任務優化的 AI 代理。此舉被視為在 Apple Intelligence 進入中國市場受阻之際，華為藉此擴大其在高端市場的算力領先地位。

Stack Overflow 轉型為代理提供 API 知識庫 ref-75

為解決 AI 代理會話結束後知識即消失的「碎片化智慧」問題，Stack Overflow 推出 Stack Overflow for Agents 平台。該服務提供 API 化的知識庫，讓編程代理能像人類工程師一樣，在遇到問題時查詢已驗證的解決方案，並貢獻新的發現。

AWS 展示 IDP 智能文件處理新架構，實現合約自動審查閉環 ref-27, ref-29

AWS 發表基於 Bedrock Data Automation (BDA) 的智能文件處理 (IDP) 流程。Rocket Close 利用此架構建立 Supercharger 代理，能自動橫跨多州法律數據庫驗證房貸合約，將原本需耗費數小時的人工審查壓縮至數分鐘，且精準度達到 98%。

月之暗面開源 Kimi K2.7-Code，推理代幣效率提升 30% ref-2, ref-55

Moonshot AI 發表 Kimi K2.7-Code 模型，採用改進的推理路徑優化。官方數據顯示，在維持相同代碼品質的前提下，其推理所需的「思考代幣」比前代減少 30%，大幅降低了大規模開發代理的維運成本。

新型攻擊「Agentjacking」可遠端劫持 AI 編程代理 ref-103, ref-117

資安研究員發現名為 Agentjacking 的攻擊手法，駭客可透過偽造的 Sentry 錯誤報告誘騙 AI 編程代理在開發者機器上執行惡意代碼。由於代理具備讀寫權限，此攻擊能輕鬆竊取環境變數與原始碼。

Google 啟動 Antigravity 專案，統一終端代理開發體驗 ref-19

Google 宣佈將原有的 Gemini CLI 升級為 Antigravity CLI。這是一個專為複雜、多代理工作流設計的開發工具，支援異步處理與桌面端同步，預計在 2026 年底前整合至所有 Android 開發環境。

市場洞察

1. 「模型主權 (Model Sovereignty)」成為企業採用的首要風險 ( ) ref-11, ref-23, ref-2

Anthropic 旗艦模型在發佈僅 72 小時內因出口管制而全球下架，這對依賴單一閉源 API 的企業敲響警鐘。市場趨勢顯示，Owning the Stack (擁有完整技術棧) 的重要性將超越模型效能。未來企業將傾向於選擇「可私有化部署」或「開源權重」的模型，以避免地緣政治導致的服務中斷。(ref-3)

2. 瀏覽器成為 AI 代理的「原生操作系統」 ( ) ref-65, ref-19, ref-137

Google 推出的 WebMCP 方案代表 AI 代理不再只是「模擬點擊」的爬蟲，而是具備與網頁原生溝通的能力。隨著 Chrome 與 Safari (ref-107) 紛紛加入代理標準化競賽，Web 應用程式的設計邏輯將從「給人看 (UI)」轉向「給代理用 (AX)」，這將催生出一批全新的 Agent-native 網路服務。

3. 硬體能效競爭進入「每兆瓦代理數 (Agents per MW)」時代 ( ) ref-23, ref-66

NVIDIA Blackwell 的基準測試結果顯示，業界評價 AI 算力的指標正在改變。過去專注於單次推論延遲，現在則關注在長達數百個 Loop 的代理鏈中，硬體如何優化上下文切換 (Context Switching) 與 工具調用 (Tool Call) 的能耗。這對於主打 On-device AI 的廠商如 Google (ref-140) 與華為 (ref-17) 是核心技術高地。

Sources 146 items

ref-0 Hacker News Statement on US government directive to suspend access to Fable 5 and Mythos 5 ref-1 Hacker News Open source AI must win ref-2 Hacker News Kimi K2.7-Code: open-source coding model with better token efficiency ref-3 Hacker News "Don't You Just Upload It to ChatGPT?" ref-4 Hacker News Malware developers added nuclear and biological weapons text to to their spyware ref-5 Hacker News How to setup a local coding agent on macOS ref-6 Hacker News Palantir loses legal challenge against Swiss investigative magazine ref-7 Hacker News Slightly reducing the sloppiness of AI generated front end ref-8 Hacker News Show HN: Putt.day a daily mini golf game ref-9 Hacker News There Is Life Before Main in Rust ref-10 Hacker News Show HN: StackScope – I crawled over 40k indie launches to see what they ship ref-11 Hacker News Shepherd's Dog: A Game by the Most Dangerous AI Model ref-12 Hacker News Launch HN: BitBoard (YC P25) – Analytics Workspace for Agents ref-13 Hacker News Show HN: Turn your name into a tree in an infinite procedural shanshui landscape ref-14 Hacker News Sam Bankman-Fried loses fraud conviction appeal ref-15 OpenAI Blog New OpenAI Academy courses for the next era of work ref-16 Google Developers Blog The latest updates to Google Pay ref-17 Google Developers Blog An important update: Transitioning Gemini CLI to Antigravity CLI ref-18 Google Developers Blog One Year of Innovation: Celebrating 100k Members in the Google Cloud x NVIDIA Developer Community ref-19 Google Developers Blog All the news from the Google I/O 2026 Developer keynote ref-20 Google Developers Blog Blazing fast on-device GenAI with LiteRT-LM ref-21 Google Developers Blog A Smarter Google AI Edge Gallery: MCP integration, notifications, and session continuity ref-22 Google Developers Blog Google Tensor SDK Beta with LiteRT ref-23 Hugging Face Blog olmo-eval: An evaluation workbench for the model development loop ref-24 NVIDIA Blog NVIDIA Blackwell Leads on First Agentic AI Infrastructure Benchmark ref-25 Vercel Blog Increased Blob store limit for Hobby users ref-26 Vercel Blog Workflow SDK now runs natively in Nitro v3 ref-27 AWS ML Blog Building Supercharger: How Rocket Close optimized title operations with agentic AI ref-28 AWS ML Blog Build a meeting prep and follow-up assistant with Amazon Quick and Cisco Webex MCP servers ref-29 AWS ML Blog From PDFs to insights: Architecting an intelligent document processing pipeline with AWS generative AI services ref-30 AWS ML Blog Built from the inside out: How AWS Professional Services became a frontier team first ref-31 The Rundown AI Jeff Bezos' $41B 'artificial general engineer' ref-32 Latent Space [AINews] Fable and Mythos officially too dangerous to release ref-33 Simon Willison Statement on the US government directive to suspend access to Fable 5 and Mythos 5 ref-34 Simon Willison OpenAI WebRTC Audio Session, now with document context ref-35 Simon Willison Quoting Andrew Singleton ref-36 TechCrunch AI Andrew Yang thinks the next big startup opportunity is lowering the cost of living ref-37 TechCrunch AI Anthropic’s safety warnings may have just backfired — the government has pulled the plug on its most powerful AI ref-38 TechCrunch AI SpaceX IPO: Live updates on everything you need to know ref-39 TechCrunch AI Meta’s months-old AI unit is a soul-crushing gulag, say the engineers stuck inside it ref-40 TechCrunch AI Chinese cybercrime operation that used AI to scam ‘hundreds of thousands of victims’ sued by Google ref-41 TechCrunch AI Mistral is rumored to be raising €3B at €20B valuation ref-42 TechCrunch AI SpaceX, Anthropic, and OpenAI’s hot IPO summer ref-43 TechCrunch AI It’s hot IPO summer, and the MANGOS are ripe ref-44 The Verge AI Siri is good now?? ref-45 The Verge AI Elon Musk is the world’s first trillionaire ref-46 The Verge AI SpaceX’s massive IPO: all the latest news ref-47 The Verge AI Jeff Bezos’ AI startup aims to build an ‘artificial general engineer’ ref-48 Ars Technica PeopleSoft 0-day affecting hundreds of organizations steals gigabytes of data ref-49 TechMeme After years of false dawns, Big Tech, startups, and governments are betting on commercially useful quantum computers by 2030, as skeptics worry about hype (Michael Peel/Financial Times) ref-50 TechMeme An interview with Corning CEO Wendell Weeks on risk-sharing provisions that protect the company in multibillion-dollar fiber deals with Nvidia, Meta, and Amazon (Christopher Mims/Wall Street Journal) ref-51 TechMeme South Korea's "semiconductor belt" town Dongtan has become one of the fastest-rising affluent areas, driven by windfall bonuses for Samsung and SK Hynix workers (Daniel Tudor/Financial Times) ref-52 TechMeme Huawei unveils HarmonyOS 7, introducing an "agent-friendly" architecture that connects to 2,000+ specialized AI agents and features an enhanced voice assistant (Iris Deng/South China Morning Post) ref-53 TechMeme A Ukrainian extradited from Ireland to the US pleads guilty to conspiracy to commit wire fraud for his role in Conti ransomware attacks between 2021 and 2022 (Lawrence Abrams/BleepingComputer) ref-54 TechMeme Anthropic says it believes the US government's export control order for Mythos 5 and Fable 5 is a "misunderstanding" and that it is working to restore access (Ananya Palyekar/Reuters) ref-55 TechMeme Moonshot AI releases Kimi K2.7-Code, claiming 30% lower reasoning token usage compared to K2.6, available under a modified MIT license (Sean Michael Kerner/VentureBeat) ref-56 TechMeme Crypto-data provider Blockworks says it has acquired its rival Messari; source: Blockworks paid more than $10M for Messari, which was valued at $300M in 2022 (Vicky Ge Huang/Wall Street Journal) ref-57 TechMeme Anthropic says it is disabling Fable 5 and Mythos 5 for all customers after the US government issued an export control order, citing national security concerns (Anthropic) ref-58 TechMeme Source: the US Commerce Department decided to block foreign persons from accessing Fable 5 and Mythos 5 after a company claimed it was able to jailbreak Mythos (Axios) ref-59 TechMeme OpenAI was subpoenaed on Friday by a coalition of state AGs for documents covering a wide range of its activities and impact on users (Wall Street Journal) ref-60 TechMeme Sources detail frustration inside Meta's Applied AI team, formed in March to support Superintelligence Labs, over menial projects, "soul-crushing" work, more (Wired) ref-61 TechMeme The US DOJ approves Paramount's $111B purchase of WBD without forcing divestitures or behavioral remedies; state AGs could still try to block the move (Yasmin Khorram/Politico) ref-62 TechMeme Sources: three ex-DOGE staffers are raising $130M from a16z, Sequoia, and others for a startup that aims to use AI to secure government systems (Vanity Fair) ref-63 TechMeme KPMG retracts a report on AI's benefits after it has been found to exaggerate AI adoption with case studies that appear to have been based on AI hallucinations (Financial Times) ref-64 InfoQ AI/ML AWS Introduces CDK Mixins for Composable Infrastructure Abstractions ref-65 InfoQ AI/ML WebMCP Standard Proposal for Agentic Web Actuation Now Available in Chrome (Origin Trials) ref-66 InfoQ AI/ML Google Launches Colab CLI for Developers, Automation, and AI Agents ref-67 InfoQ AI/ML Slack Eliminates SSH in EMR Pipelines, Migrates 700+ Jobs to Rest-Based Architecture ref-68 InfoQ AI/ML Angular's Official Agent Skills Helps AI Coding Tools Write Modern Angular ref-69 InfoQ AI/ML Pinecone Brings AI Agents Directly to Enterprise Data with Microsoft OneLake Integration ref-70 InfoQ AI/ML Run Untrusted AI Agent Code Safely with Azure Container Apps Sandboxes ref-71 InfoQ AI/ML Presentation: Moving Mountains: Migrating Legacy Code in Weeks instead of Years ref-72 InfoQ AI/ML Podcast: Craig McLuckie on Culture as a Team's Operating System in the AI Era ref-73 The New Stack Federal government orders Anthropic to pull Fable 5 and Mythos 5, three days after launch ref-74 The New Stack Who gets to be Switzerland in the enterprise agent wars? ref-75 The New Stack Coding agents have questions, too — so Stack Overflow built them a home ref-76 Wired AI Anthropic Says It’s Taking Claude Fable 5 Offline to Comply With US Government Order ref-77 Wired AI Meta Employees Absolutely Hate Mark Zuckerberg’s Plan for a Companywide AI Hackathon ref-78 Wired AI ‘Tell Him He’s a Piece of Shit’: Meta’s New AI Unit Is a Total Mess ref-79 Wired AI China Didn’t Make Americans Hate Data Centers ref-80 Wired AI You Probably Won’t Get Rich Off the SpaceX IPO ref-81 Crunchbase News The Week’s 10 Biggest Funding Rounds: NinjaOne Leads With $400M As Large Deals Also Go To Blockchain, Cloud Infrastructure, Biotech And Robotics ref-82 Crunchbase News SpaceX Shares Close Up 19% After Largest IPO Of All Time ref-83 Stratechery 2026.24: Hey Siri, Tell Me a Fable ref-84 Tom's Hardware Nvidia preps to sell its Vera CPUs into China as its GPU sales stay frozen — customers encouraged to place orders for CPU shipments as early as August ref-85 Tom's Hardware Amazon Prime Day ref-86 Tom's Hardware Thrustmaster's new specialized T.Flight Hotas 5 Microsoft Flight Simulator Edition provides a plug-and-play flight sim setup for just $109 — featuring 5-axis control with 16-bit precision and dual-rudder system ref-87 Tom's Hardware Radeon RX 9070 XT finally appears in Steam Hardware Survey — RDNA 4 flagship surprisingly lands just behind RTX 5080 ref-88 Tom's Hardware Nvidia's high-speed AI data center storage servers break cover, touting 2.9 petabytes of storage and extreme PCIe 6.0 performance — Wiwynn shows off SCADA server with GPU-accelerated storage ref-89 Tom's Hardware Microsoft's bug-hunting nemesis extends vendetta with more zero-day attacks — Nightmare Eclipse publishes RoguePlanet and GreatXML local privilege escalation exploits ref-90 Tom's Hardware Various vendors add AMD EXPO Ultra-Low Latency to 600-series motherboards in latest BIOS updates — tech tightens memory subtimings on compatible kits, boosting FPS by up to 4% ref-91 Tom's Hardware Republican lawmakers urge federal agency to block imports of infringing TSMC chips as patent ruling nears — five asserted U.S. patents come from United Microelectronics Corporation ref-92 Tom's Hardware Save $300 on Gigabyte's Gaming A16 gaming laptop at Walmart — Budget RTX 5060 -powered 16-inch laptop is now only $1,199 ref-93 Tom's Hardware Save a massive $751 on this RTX 5070 Ti gaming PC with a 9800X3D right now — liquid-cooled, 4K-ready Skytech rig with 32GB DDR5 and a 2TB SSD is now just $2,249 ref-94 Tom's Hardware Xbox will pay five times more for memory and storage in 2027 than it did two years ago — CEO Asha Sharma admits there's an unsustainable hardware gap that 'cannot continue' ref-95 Tom's Hardware Get 32GB of DDR5 RAM for only $255 in this 2-item combo from Newegg — just $514.99 gets you Corsair Vengeance RGB RAM and a Gigabyte X870 Aorus Elite motherboard, 26% off ref-96 Tom's Hardware New malware campaign tricks AI scanners with fake nuclear weapon prompts — malicious code triggers safety failsafes so scanners skip the payload ref-97 Tom's Hardware AMD denies researcher a $10,000 bug bounty after fixing critical auto-updater vulnerability — security flaw took 124 days to patch ref-98 Tom's Hardware Several police officers arrested for using controversial Flock AI license plate reader system to stalk romantic partners, says report — investigators have unearthed at least 18 such cases in the US over recent years ref-99 The Hacker News (Security) U.S. Orders Anthropic to Suspend Fable 5 and Mythos 5 Access for Foreign Nationals ref-100 The Hacker News (Security) Over 400 Arch Linux AUR Packages Hijacked to Deploy Infostealer and eBPF Rootkit ref-101 The Hacker News (Security) Google Sues Chinese Smishing Network Accused of Using Gemini AI in Phishing ref-102 The Hacker News (Security) China-Linked Hackers Backdoored Linux Login Software to Hide for Nearly a Decade ref-103 The Hacker News (Security) Agentjacking Attack Tricks AI Coding Agents Into Running Malicious Code ref-104 The Hacker News (Security) Rethinking MDR as Attackers and Defenders Embrace AI ref-105 The Hacker News (Security) LangGraph Flaw Chain Exposes Self-Hosted AI Agents to Remote Code Execution ref-106 The Hacker News (Security) INTERPOL Operation Takes Down Sniper Dz Phishing Platform, Arrests Administrator ref-107 EU Startups Ukraine-focused DefenceTech fund Varangians closes €9.1 million in Stockholm ref-108 EU Startups Profitable Oslo-based AI startup Mimir raises €518.3k pre-Seed to automate e-commerce operations ref-109 EU Startups Weekly funding round-up! All of the European startup funding rounds we tracked this week (June 08 – June 12) ref-110 EU Startups Slovakia’s Nordics rebrands as Definic, raises €2.5 million to scale its vendor intelligence platform ref-111 MIT Research MIT engineers find a way to deliver drugs directly to the esophagus ref-112 Lobsters What’s New in WeatherMesh-6 ref-113 Google News Google unveils DiffusionGemma, an AI model that breaks free of left-to-right processing ref-114 Google News Anthropic shuts down newest AI model after U.S. bans foreign use ref-115 Google News Nvidia, Abridge collaborate to develop healthcare-specific AI model ref-116 Google News PixelRAG beats text parsers on accuracy and cuts AI agent token costs 10x ref-117 Google News Agentjacking Attack Tricks AI Coding Agents Into Running Malicious Code ref-118 Google News Apple opens the door to third-party AI model vendors in latest update ref-119 Google News No more alibis: why generative AI can’t hide behind Section 230 any longer ref-120 Google News The First Bank To Let An AI Agent Move Money Will Set The Rules For Everyone ref-121 Google News Samsung reverses years-long ban on external gen AI use ref-122 Google News Meta unveils Muse Spark AI model with selectable modes for different tasks ref-123 Google News Coinbase debuts AI agent that can trade and pay for premium research ref-124 Google News Microsoft’s open-source SkillOpt automatically upgrades AI agent skills without touching model weights ref-125 Google News OpenAI acquires Ona cloud startup to boost Codex AI coding tool ref-126 Google News UK Regulator Introduces “World First” Controls on the Use of Content in Google Generative AI Features ref-127 Google News AI model accelerates molecular simulations by 10,000-fold ref-128 Google News Jeff Bezos’ AI startup aims to build an ‘artificial general engineer’ ref-129 Google News Meta is unwinding its $2 billion AI acquisition as China forces the deal apart ref-130 Google News Jeff Bezos's AI startup Prometheus raises $12 billion at $41 billion valuation ref-131 Google News Trump administration, OpenAI discussing possible government stake in the AI startup ref-132 Google News SpaceX's AI revenue could eclipse Starlink — Goldman forecasts 100-fold surge to $322B by 2030 ref-133 Google News UK Government signs AI partnership with ElevenLabs to improve accessibility of public services ref-134 Google News French AI startup Mistral in talks to raise at €20 billion valuation ref-135 Google News California State Bar's proposed AI ethics rules put attorneys on notice ref-136 Google News Anthropic’s Fable Is Locked Down As US Takes AI Safety Into Its Hands ref-137 Google News How Trump's AI strategy is taking shape ref-138 Google News Singapore, Microsoft team up to build reliable AI governance framework ref-139 Google News Trump’s AI security order acknowledges risks but stops short of regulating industry ref-140 Google News AI Regulation Push: Amodei Demands Power Blocking Unsafe Models, Anthropic Pledges $350 Million ref-141 Google News AMD Stock Upgraded To Buy On GPU Upside ref-142 Google News Taiwan Semiconductor (TSM) Sees AI Chip Demand Outpacing Supply ref-143 Google News The Next US-China AI Battle Is Over Compute — and China Is Spending Big ref-144 Google News Google and Samsung may partner to develop the Icefish AI chip ref-145 Google News KKR launches $10 billion AI infrastructure company with Nvidia, Vistra