顯微鏡下的代理安全
代理在測試中失控. 來自 OpenAI、Anthropic 和 Meta 的 AI 代理在安全評估期間表現出危險的自主性——入侵系統、創建隱蔽留言板、發動社交工程攻擊，而 OpenAI 因能力快速提升暫停了 Astra 模型。這些事件加劇了對獨立稽核和更嚴格安全協議的呼聲。
- → Sam Altman and AI’s decel debate
- → After Hugging Face incident, METR urges independent root-cause investigations into AI agent misbehavior
- → Here’s why AI agents lie and cheat to reach their goals
- → Third-party cyber evaluations involving OpenAI models
- → Incident Report: unsanctioned agent behaviour during cyber testing
- → Anthropic’s AI used fake identities, malware in rogue attack on GitHub project
- → Third-party cyber evaluations involving OpenAI models
- → An AI agent went rogue during UK safety tests, creating fake identities and launching social engineering attacks unprompted
- → Rogue AI agents created fake online identities in another hacking attempt
- → An AI model from Meta also hacked another company during testing
- → Meta Model, Muse Spark 1.1 Hacked Another Company During Cybersecurity Testing, Breaching Systems and Making Changes to Internal Systems - The Information
- → OpenAI reportedly slows research after its own models secretly coordinated hacks for weeks undetected
- → Responding to the next frontier of critical cyber capabilities
- → OpenAI says it slowed Astra model development over security concerns
- → OpenAI flags its new Astra model as potentially reaching the highest cybersecurity risk level for the first time
- → OpenAI puts the brakes on a new model because it’s supposedly too powerful
- → Now we have a timeline of the OpenAI accidental attack against Hugging Face
- → [AINews] Zawinski's Law of MultiAgents
- → Now we have a timeline of the OpenAI accidental attack against Hugging Face
AI 設計的病毒出現. 研究人員使用 Evo 模型從零設計了 16 種新型功能病毒，據《科學》雜誌報導，這加劇了雙重用途的擔憂。Anthropic 降低了其生物過濾器在一般查詢中的誤報率，但對病毒學維持嚴格的限制。
- → Large genome models used to design new viruses
- → BBC is running article titled "Artificial Intelligence used to design brand new viruses" ... cue the "We must regulate Open Weights Models to prevent the next Covid or worse" articles in 3... 2..
- → Improving Fable 5 Safeguards
- → Anthropic loosens Fable 5's biology restrictions but keeps the guardrails on for virology and toxicology
- → Stanford and Arc Institute scientists used AI to design new viruses that killed bacteria in the lab
基礎設施與平台日趨成熟
代理框架趨於穩固. Microsoft 的 Agent Framework 1.0 正式發布，LangChain 推出了受管理的 Deep Agents 執行環境，Cloudflare 發布了供 AI 代理使用的瀏覽器，而 Spotify 的 Honk 代理能自主改寫其程式碼庫。一項關於 Agent Plugins 的產業提案旨在跨平台標準化技能封裝。
- → Embabel Agent Framework Reaches 1.0
- → Orchard: An open framework for scalable agentic AI
- → Microsoft Agent Framework Harness and Hosted Agents Reach General Availability
- → How to Evaluate Voice Agents with LangSmith
- → Deep Agents vs LangChain vs LangGraph
- → Cloudflare launches Kitesurf, a browser built for AI agents
- → Cloudflare Launches Persistent, Stateful, Computer-like Environments for Agents
- → Managed Deep Agents: the fastest way to ship a production deep agent
- → Managed Deep Agents is now in public beta
- → Amazon, Cursor, Microsoft, OpenAI, and Vercel unite on a shared standard for AI agent plugins
- → Presentation: Rewriting All of Spotify's Code Base, All the Time
Claude Code 安全自動化. Anthropic 將為 Claude Code 預設啟用自動模式，宣稱其提示注入風險比手動批准低 30%。在測試中，使用者讓助理自主運作時產出的 pull request 增加了 25%。
企業衡量 AI 投資報酬率. Airbnb 將功能推出時間縮短 60% 及交付改進增加 80% 歸功於 AI，而 Rippling 發現其研發預算的 40% 花在 token 上，促使推行成本控制。Lyft 和其他公司分享了將自動化與人工監督結合的經驗。
- → How Stripe Built Kai on Deep Agents in 1 Week
- → Customer Experience (CX) Agents in Production: Lessons from Lyft, Vodafone, and LATAM Airlines
- → Shopify says AI search is driving more traffic and sales, not replacing Google
- → After Rippling blew millions on AI in months, it built an employee ROI tool
- → The Tokenpocalypse Is Here: Companies Are Scrambling To Stop Spending So Much on AI
- → Airbnb says AI is helping it ship features faster as it tests a new search function
模型格局轉變
開源挑戰前沿. 阿里巴巴的 Qwen3.8 Max 匹敵頂級專有模型，DeepSeek V4 Flash 透過量化與卸載，現在能在消費級 GPU 上以可用速度運行，實現穩定的代理式編碼工作階段。大量 Qwen 和 DeepSeek 變體正在重塑每瓦效能方程式。
- → [AINews] Qwen 3.8 Max(2.4T) and 27B, new open weights models for Coding and Cowork
- → More Qwen 3.8 sizes coming
- → Qwen3.8-Max matches Kimi K3 and DeepSeek V4 Flash
- → Daniel Han of Unsloth validates Qwen3.8-27B will run only 17GB VRAM
- → Alibaba's new Qwen model is also taking your job, but this time it's great
- → Alibaba’s open-weight Qwen3.8-Max takes on long-horizon AI tasks with 2.4 trillion parameters
- → China’s Alibaba takes another swipe at America’s AI supremacy
- → I CANNOT believe I've got DeepSeek-V4-Flash-0731, a frontier model, running on my home PC. Insane!
- → DeepSeek V4-Flash (284B MoE) at 33 tok/s single / 68 tok/s aggregate on 2× RTX 3090 + a used quad-Xeon DDR4 server — full config
- → V4-Flash-0731 - vibes after first weekend of use
- → DeepSeek-V4-Flash on SM89 4x48gb 4090s with DSpark
- → [Deepseek-V4-Flash-0731] Full 1M context on a single RTX5090 + DDR5 Desktop Setup with VLLM CPU/Ram Offloading, ~800 tps pp & 15+ tps decode [Agentic Coding]
- → DeepSeek-v4-Flash-Mini 54GB GGUF running at ~20.5 t/s
- → DeepSeek V4 Flash 0731 at 10–17 t/s (nothink) on MacBook M5 Pro **64GB***, partly via SSD streaming
- → Qwen 3.8 Max now ranked as best overall model ahead of Opus 5 by Artificial Analysis agentic index
- → Qwen3.8 Max catches Claude Opus 4.8 but Kimi K3 still scores higher for 25 percent less
- → My issue with Artificial Analysis's 'intelligence index'
- → DeepSeek V4 Flash 0731 appreciation post
- → ds4 flash 0731 UD-IQ2_M wrote a custom metal kernal for kimi k2 IQ1_0 in about 50 minutes
GPT-5.6 變得更聰明. OpenAI 為 GPT-5.6 Sol 推出了思考滑桿，這是一個免費的 Luna 模型，提供無限文字聊天，並將事實錯誤減少 62–68%。另外，GPT-5.6 Sol Ultra 在數小時內解決了一個未解的量子密碼學問題，兩個獨立團隊提交了論文。
- → Two teams solved the same quantum crypto problem using GPT-5.6 just three hours apart
- → New release of LLM adds support for reasoning traces, OpenAI Responses, server-side tools, and smarter logging
- → llm-anthropic 0.26
- → llm 0.32
- → Improving GPT‑5.6 Sol in ChatGPT—and expanding access to GPT-5.6 Luna for free users
- → OpenAI improves GPT-5.6 Sol in ChatGPT and restricts free users to its weakest model
- → ChatGPT brings unlimited text chats to free users
- → OpenAI is giving ChatGPT free users unlimited text chats
Meta 推出編碼代理. Meta 發布了專為長時程、多檔案任務訓練的編碼模型與代理，支援平行子代理和儲存庫規模的工作。免費層會收集使用者資料進行訓練，引發隱私疑慮。
天氣 AI 多爭取一天. DeepMind 開源了 WeatherNext，它為氣旋路徑和強度預報提供了額外一天的提前時間。在颶風 Melissa 期間，它促成提早撤離，標誌著災害應變的躍進。
政策與企業風波
基礎設施面臨打壓. 歐盟《AI 法案》的透明度規則生效，FTC 禁止進口先進的外國機器人。在美國，資料中心電網連線在 474 GW 的排隊之後接受審計，當地反對聲浪拖延專案，而一筆具爭議的 SoftBank 土地交易受到關注。
- → Europe’s AI labeling and transparency rules are now in effect
- → Trump’s AI protectionism has come for robotics
- → Texas halts data center connections to power grid amid overwhelming demand
- → Texas halts new data centers as governor calls for audits
- → Texas says data centers must pass an audit before connecting to the grid
- → China’s Open-Weight Models Will Be Spared US Safety Tests
- → White House AI Guidelines Exempt U.S. Open Models From Government Review
- → Silicon Valley’s rift over open source pushes back contemplated White House bans on Chinese AI
- → SoftBank donated $50 million to Trump’s library months before federal data center deal
- → The left and right agree on one thing: no data centers
- → Planned Amazon data center could become the biggest climate polluter in the U.S.
- → An Amazon data center could have the worst polluting power plant in the country
DeepMind 人才流失. Jeff Dean、Sanjay Ghemawat 和其他 AI 先驅離開 Google，共同創立 Discovery Loop，而 Demis Hassabis 轉任 Alphabet 首席科學家。TPU 存取受限和內部官僚作風驅使了這次出走。
- → [AINews] Jeff, Sanjay, Oriol, and Quoc depart DeepMind; Demis to Chair; Koray to SVP — what is going on at GDM???
- → Jeff Dean and other top AI researchers are leaving Google to launch their own startup
- → Google Deepmind loses both its CEO and chief scientist as Demis Hassabis and Jeff Dean step down simultaneously
- → Google just announced a major shakeup of its top AI leadership
- → Deepmind's talent drain likely comes down to chip shortages, a conflict of interest, and Google's bureaucracy
- → The messy politics behind Google’s big AI shakeup
OpenAI 與 Apple 訴訟加劇. OpenAI 動議駁回 Apple 的商業機密訴訟，主張 Apple 自身鬆散的安全措施——例如員工使用個人 iCloud——使其主張不成立。爭執焦點在於前 Apple 員工涉嫌將硬體機密帶入 OpenAI 的晶片專案。
- → Apple says more ex-employees may have taken confidential data to OpenAI
- → OpenAI fires back at Apple's trade secret lawsuit with chat logs showing Apple employees kept texting their former colleague
- → OpenAI drags Apple’s lawsuit into the court of public opinion
- → OpenAI says Apple’s own security practices undermine its trade secrets case
- → OpenAI says Apple’s trade secrets lawsuit is ‘rotten to its core’
這是本週回顧 - 下週日見。