AiX 環球 AI 情報 Global AI Intelligence

第 35 期 2026 年 11 月 1 日（星期日） 07:30 出刊 10 條情報 · 15 原文來源連結 RSS 過往期數

今日十條 · 2026 年 11 月 1 日算法的權限正在被逐層收回：代理打穿供應商、瀏覽器攔下收據、政策提前命名工種

今日主編判斷

今日四條線指向同一件事：誰有權讓算法做什麼。國務院辦公廳首份回應人工智能就業效應的文件，把失業緩衝期、應用前告知寫成可執行義務；北京以三千萬元補貼模型首方案；美國聯邦貿易委員會要九家企業交代算法如何定價；Chainalysis 承認被攻擊型代理技術攻破。權限與紀錄已取代跑分。Four threads converge on one question: who is authorised to let an algorithm do what. China issued its first State Council document on AI's employment effects, turning redundancy buffers and advance notice into executable duties; Beijing attached up to 30 million yuan to a first deployment; the US FTC ordered nine firms to explain how algorithms set prices; and Chainalysis admitted an intrusion using offensive agentic techniques. Permission and record have replaced leaderboard scores.

今日十條

本欄由 AIX Society 編輯部整理公開資料後撰稿，每條均附原始來源連結；事實與觀點分列，觀點屬本會判斷。

按受眾篩選

國務院辦公廳首份人工智能就業文件出台，六方面十八項任務寫入失業緩衝期與應用前告知China's State Council issues its first AI employment document, with 18 tasks including redundancy buffers and advance notice

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners人力資源HR法務合規Legal & compliance

事件: 國務院辦公廳印發《關於適應人工智能技術發展助力構建就業友好型發展方式的意見》，設十八項任務，提出失業緩衝期、應用前告知受影響員工，並對高衝擊項目作就業風險評估。China's State Council General Office issued a document on building employment-friendly development amid AI adoption. Its 18 tasks across six areas encourage redundancy buffers, require advance notice before large-scale deployment, and mandate employment risk assessments for high-impact projects.
背後: 這是國務院辦公廳首次專門就人工智能就業效應印發的頂層設計，出台時正值各地壓縮招聘，把社會成本由企業內部轉為公共議題。This is the first top-level design issued specifically on AI's employment effects by the State Council General Office. It lands amid hiring freezes and headcount cuts across regions, moving social cost from a firm's internal matter to a public question.

顧問觀點

我們認為文件真正改變的是舉證責任而非補貼：企業今後要證明部署前做過評估與告知。把人力規劃與技術採購放在同一張決策桌上，比事後補救便宜得多。We read the change as a shift in the burden of process evidence rather than subsidies: firms must show an assessment and a notice preceded deployment. Putting workforce planning beside technology procurement costs far less than retrofitting.

對日常工作的影響: 人力資源主管要把就業風險評估併入技術立項表，並與資訊部門共用一份受影響崗位清單，否則首批被追問的就是招聘與轉崗紀錄。HR leads should fold the employment risk assessment into the technology intake form and share one list of affected roles with IT, or the first documents demanded will be hiring and redeployment records.
如何改善: 本週挑一項正在推進的流程自動化，補寫一頁受影響崗位、告知時點與轉崗安排並存檔備查。This week, take one live process automation and add a one-page record of affected roles, notice timing and redeployment plans, then file it.

來源中國經營報（2026-10-31）↗中國新聞社（2026-10-31）↗

OpenAI 把本輪估值目標由一點四萬億美元下調至一點二五萬億，募資額減至二百億OpenAI cuts its round to 20 billion dollars at a 1.25 trillion valuation, down from 30 billion at 1.4 trillion

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners資訊科技IT法務合規Legal & compliance

事件: 《金融時報》引述知情人士，OpenAI 本輪募資目標由三百億減至約二百億美元，目標估值由一點四萬億降至一點二五萬億，週五為承諾截止日。Citing people familiar with the matter, the Financial Times reported that OpenAI cut its latest round from 30 billion to about 20 billion dollars and its valuation target from 1.4 trillion to 1.25 trillion. A Friday deadline was set for commitments, and the first close is expected below 20 billion after some investors reduced allocations.
背後: 納斯達克指數單日跌百分之二點九，較十月六日高位回落百分之七點八；微軟、谷歌與 Meta 同週籌資，資金轉趨審慎。Around the report the Nasdaq fell 2.9 per cent in a day, down 7.8 per cent from its 6 October record. Microsoft, Google and Meta all raised capital the same week as investors grew warier of compute commitments.

顧問觀點

我們認為這是訂價而非崩塌。同一批投資者要同時認購兩家現金流為負的實驗室，可分配資金本身有上限。企業續約籌碼變多，但要留意供應商為鎖定收入而加推長約。We read this as repricing rather than collapse. The same investor pool is asked to fund two cash-negative labs, so available capital has a ceiling. Buyers gain leverage at renewal, but should watch vendors pushing long contracts to lock in revenue.

對日常工作的影響: 財務總監與資訊主管要重看手上的多年期算力或模型合約，評估違約與漲價風險，並在採購條款補上用量下限與退出機制。CFOs and IT leads should revisit multi-year compute or model contracts, weigh default and repricing risk, and add usage floors and exit clauses to procurement terms.
如何改善: 本週把現有最長一份人工智能相關合約抽出來，列出解約條款與用量承諾，向管理層寫三點風險摘要。This week, pull your longest AI-related contract, list its termination clauses and volume commitments, and give management a three-point risk note.

來源Reuters（2026-10-30）↗中新經緯（2026-10-31）↗

美國聯邦貿易委員會向九家企業發出六(b)調查令，追查算法如何決定價格與零工報酬The US FTC issues 6(b) orders to nine firms over algorithmic pricing and gig pay

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners市場推廣Marketing法務合規Legal & compliance資訊科技IT

事件: 聯邦貿易委員會以二比零通過，向達美、聯合、捷藍、Lyft、Uber、DoorDash、Instacart、Target 與沃爾瑪發第六(b)條調查令，查定價是否隨位置與收入變化。The Federal Trade Commission voted 2-0 to issue Section 6(b) orders to Delta, United, JetBlue, Lyft, Uber, DoorDash, Instacart, Target and Walmart, asking whether prices vary with a shopper's location, income and browsing history.
背後: 本輪由中介機構轉向直接定價的企業。企業須四十五日內回覆，拒絕可被訴至聯邦法院；披露建議意見期延至十一月二十一日。This round moves from intermediaries to the firms that set prices. Companies must make contact within 14 days and respond within 45, with refusal taken to federal court. Comments on a disclosure proposal now close on 21 November.

顧問觀點

我們認為監管的落點在模型的輸入特徵，而非模型本身。只要特徵清單存在可反推收入、住址或健康狀況的代理變數，即使沒有直接使用敏感屬性，解釋成本依然極高。We see the target as the model's input features rather than the model. Where a proxy for income, address or health exists in the feature list, the cost of explanation stays high even without direct use of a protected attribute.

對日常工作的影響: 市場與法務主管要為每條差別定價規則寫下商業理由，由資訊部門輸出特徵清單；客服前線需要一句可對外解釋的統一口徑。Marketing and legal leads must record a commercial rationale for each differential pricing rule, IT must export the feature list, and frontline staff need one approved line for customers.
如何改善: 本週選一個在用的動態定價模型，列出前二十個輸入特徵，逐項標註能否反推個人敏感狀況。This week, take one live dynamic pricing model, list its top 20 input features, and mark each for whether it can infer a sensitive personal circumstance.

來源Finks AI（2026-10-30）↗

Chainalysis 承認遭「攻擊型代理技術」入侵，客戶姓名、電郵與職務標籤被讀取Chainalysis confirms an intrusion using offensive agentic techniques, exposing customer names, emails and role labels

影響力✓ 已核實Verified已在發生Happening now資訊科技IT企業決策者Business owners法務合規Legal & compliance

事件: 區塊鏈分析公司 Chainalysis 於十月三十日通告，本月偵測到外部攻擊者以攻擊型代理技術進入其系統，部分客戶資料被讀取，含用戶編號、全名、電郵與職務標籤。Chainalysis published a security notice on 30 October saying it detected an external actor using offensive agentic techniques to gain unauthorised access to its systems and internal tools. Non-sensitive customer data was accessed, including user IDs, full names, email addresses and organisational role labels.
背後: 客戶多為執法機構、政府與金融機構。公司稱未發現調查內容、密碼、多因素密鑰或工作階段權杖外洩，並表示與近日幣安被盜無關。Its customers are largely law enforcement, government and financial institutions. The firm said it found no access to investigation content, passwords, multi-factor secrets or session tokens, and that the incident is unrelated to the recent Binance theft.

顧問觀點

我們認為關鍵不在有沒有用人工智能，而在機器速度把偵察、橫向移動與清理痕跡壓進同一個時間窗，人類值班表接不上。防線要建在權限邊界，不是告警數量。We see the change as machine speed compressing reconnaissance, lateral movement and cleanup into one window no human rota can match. Build the defence at the permission boundary, not on alert volume.

對日常工作的影響: 資訊主管要把內部工具與對外平台分區，並假定供應商外洩必附職務資訊；客服團隊遇自稱供應商來電，須改用既有渠道回撥。IT leads should segment internal tools from customer-facing platforms and assume vendor breaches carry role data; service teams should call back on a known channel when a self-described supplier gets in touch.
如何改善: 本週列出內部工具的公網暴露面，關閉三個以上不必要入口，並抽查一次職務資料的最小權限。This week, list the public exposure of internal tools, close at least three unnecessary entry points, and sample one least-privilege check on role data.

來源The Hacker News（2026-10-30）↗InYourGeek（2026-10-31）↗

OpenAI、谷歌、微軟與 Anthropic 同認一個 robots.txt 訓練退出令牌，生效時間各異OpenAI, Google, Microsoft and Anthropic back one robots.txt training opt-out, on different timelines

影響力✓ 已核實Verified6–12 個月6–12 months市場推廣Marketing法務合規Legal & compliance資訊科技IT

事件: 四家實驗室十月三十日承認 robots.txt 共用令牌 __usage_ai-training，網站一行即可對四家爬蟲表達不參與訓練。令牌出自 Anthropic 的 IETF 草案；微軟訂二〇二七年二月、谷歌二〇二七年、OpenAI 稱數月。The four labs said on 30 October they will honour a shared robots.txt token, __usage_ai-training, letting a site state one training preference for all four crawlers. The token comes from an Anthropic individual IETF draft, with Microsoft targeting February 2027, Google 2027 and OpenAI the coming months.
背後: 令牌只表達訓練偏好，不是抓取指令，也不追溯已收集內容；屬個人草案，IETF 未背書，遵守全憑自願，網站須保留原有規則。The token expresses a training preference only, is not a crawling directive and does not reach content already collected. It is an individual draft without IETF endorsement, honoured voluntarily; sites must keep per-vendor rules.

顧問觀點

我們認為它降低成本多於解決授權：過去要逐一舉證自己拒絕過，今後一行配置就是證據鏈起點。內容授權與定價仍要回到合約。We see lower cost rather than settled licensing: instead of proving a refusal vendor by vendor, one configuration line starts the evidence trail. Licensing and pricing still return to contracts.

對日常工作的影響: 市場主管要同步更新全站 robots.txt，並與搜尋曝光負責人確認不誤傷索引；法務把該行偏好列為內容授權談判前置條件。Marketing leads should update robots.txt estate-wide and confirm with whoever owns search visibility that indexing is unaffected, while legal lists the preference as a precondition in content licensing talks.
如何改善: 本週一次過審核全站 robots.txt，加入共用令牌行，並在改動記錄寫明保留的搜尋規則。This week, audit every robots.txt in one pass, add the shared token line, and log which search rules were retained.

來源Emergent（2026-10-31）↗MACCUS（2026-10-31）↗

北京市出台智能體措施，「模型首方案」最高補貼三千萬元、目標算力十萬 PBeijing offers up to 30 million yuan for first deployments of agent models and targets 100,000 P of compute

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners資訊科技IT

事件: 北京市於十月三十一日介紹智能體措施，對模型首方案最高補貼三千萬元；北京已開發智能體逾一千二百個，算力目標十萬 P。Beijing outlined measures on 31 October to accelerate agents, supporting AI-native super software and AI operating systems with up to 30 million yuan for a first model-first solution. The city says it has built over 1,200 agents and pooled more than five billion industry data records, targeting 100,000 P of compute.
背後: 補貼指向首次部署而非研發，等於替企業墊付試錯成本。政策詞彙由大模型轉向智能體與場景，反映能力已足、落地不足的判斷。The subsidy targets first deployments rather than research, effectively prepaying trial and error. Vocabulary has shifted from large models to agents and use cases, an acknowledgement that capability is not the binding constraint.

顧問觀點

我們認為補貼會催生一批示範項目，但驗收若仍以「上線幾個智能體」計，就會重演過去資訊化項目的數字遊戲。該看的是每月調用次數與替代的工序。We expect a wave of showcase projects. If acceptance is still counted in agents launched, the numbers game of earlier IT programmes repeats. What matters is monthly invocations and process steps replaced.

對日常工作的影響: 資訊主管要把補貼申報與內部工單打通，讓每個示範代理對應一道可量度工序；財務須把補貼、硬件與運維合成三年現金流。IT leads should connect subsidy applications to internal ticketing so each showcased agent maps to one measurable process, while finance folds subsidy, hardware and running costs into a three-year cash flow.
如何改善: 本週選一道可量度工序，寫成兩頁首方案申請草案，附上前後處理時間與人力投入對比。This week, choose one measurable process, draft a two-page first-deployment application, and attach before-and-after handling times and staffing.

來源機械姬安娜（引北京市經信局）（2026-10-31）↗

Anthropic 調整 Claude 訂閱：Max 20x 更名 Ultra 並取消五小時限額，每週額度不變Anthropic renames Max 20x to Ultra and drops the five-hour cap, leaving weekly limits unchanged

影響力✓ 已核實Verified已在發生Happening now資訊科技IT人力資源HR

事件: Anthropic 於十月三十一日宣布，Max 5x 更名 Max（一百美元）、Max 20x 更名 Ultra（二百美元）；已驗證 Ultra 帳戶取消五小時限制，週額度仍為 Pro 十倍。Anthropic announced on 31 October that Max 5x becomes Max at 100 dollars a month and Max 20x becomes Ultra at 200 dollars. Verified Ultra accounts lose the five-hour cap and can work continuously within a weekly allowance still ten times Pro.
背後: 官方稱原名 20x 令人誤以為兩項限額同為 Pro 的二十倍，負責人已致歉。調整在 OpenAI 把用量翻倍之後一日。The stated reason is that the 20x name suggested both limits were 20 times Pro, and the subscription lead has apologised. The change follows OpenAI doubling allowances on several tiers a day earlier.

顧問觀點

我們認為這是把限額由總量改成時段，真正影響交付的是排隊而非總額。企業該問的不是月費多少，而是長任務會不會在下午被切斷。This converts a cap on volume into a cap on cadence; queueing, not the total, affects delivery. The enterprise question is not the monthly fee but whether a long task gets cut off mid-afternoon.

對日常工作的影響: 資訊主管要以連續可運行時長而非帳單金額評估開發工具，並為代理長任務設計分段存檔；人力資源要更新工具津貼檔位對照表。IT leads should judge developer tooling by continuous runtime rather than bill size and design checkpointing for long agent tasks, while HR updates allowance tiers to the new names.
如何改善: 本週量度一次最長自動化任務的實際耗時，與各檔位限額曲線對照，再決定是否升級。This week, measure the actual duration of your longest automated task against each tier's allowance curve before deciding on an upgrade.

來源機器之心（2026-10-31）↗IT之家（2026-10-31）↗

韓國個人資訊保護委員會修訂安全措施標準，第四、五、六、八條於十月三十一日生效South Korea's PIPC amendments to personal data security standards take effect on 31 October

影響力✓ 已核實Verified已在發生Happening now法務合規Legal & compliance資訊科技IT人力資源HR

事件: 韓國個人資訊保護委員會修訂的《個人資訊安全性確保措施標準》第四、五、六及第八條於十月三十一日生效，第四條要求機構制定並實施安全措施。Articles 4, 5, 6 and 8 of the Korean Personal Information Protection Commission's amendment to the Standards for Measures to Ensure the Safety of Personal Information took effect on 31 October 2026, with Article 4 requiring organisations to establish and implement security measures. Consultation began in July 2025, with earlier articles in force from October 2025.
背後: 南韓近期接連出現電訊、金融與建築業資料外洩，監管壓力由事後處罰轉向事前證明措施到位，修訂把義務落到可核查的條文。South Korea has seen successive breaches across telecoms, finance and construction. Regulatory pressure is shifting from after-the-fact penalties to proof that controls exist in advance, and the amendment makes obligations auditable.

顧問觀點

我們認為示範意義大於條文本身：它把「有沒有做」與「做得好不好」分開，前者成為行政處罰門檻。網絡安全規則亦被併入個資標準，兩套追責路徑合流。We read the signal as larger than the clauses: it separates whether a control exists from how good it is, making the former the penalty threshold. Cyber rules are folded into the privacy standard, merging two accountability routes.

對日常工作的影響: 法務與資訊主管要重做資料處理紀錄，標明每類個資的加密、存取與日誌措施，並訂一份共同的事故分級表以統一簽核。Legal and IT leads should rebuild the processing record, marking encryption, access and logging for each data category, and agree one incident severity table so approvals stop stalling.
如何改善: 本週列出一個存放韓國客戶資料的系統，補齊存取日誌與加密方式兩項紀錄，交合規負責人簽核。This week, take one system holding Korean customer data, complete its access log and encryption entries, and have compliance sign off.

來源Digital Policy Alert（2026-10-31）↗

谷歌在新加坡設研究中心並與兩大醫療集團合作，以開源模型在院內運行臨床助手Google opens a Singapore research centre and partners two health clusters on in-hospital clinical assistants

影響力✓ 已核實Verified6–12 個月6–12 months資訊科技IT企業決策者Business owners法務合規Legal & compliance

事件: 谷歌於十月二十九日在新加坡宣布設研究中心，並與國立大學醫學組織合作評估整合指南與文獻的臨床助手，另與國立健保集團以人工智能分析人群健康風險。At its first healthcare summit in Singapore on 29 October, Google announced a local research centre and two partnerships: with NUHS to evaluate a clinical assistant consolidating guidelines and literature, and with NHG to model population health risk and push prevention advice through NHG HealthBot.
背後: 合作採用可部署於醫院自有系統的開源模型 MedGemma，病歷與影像毋須上傳谷歌伺服器。新加坡五月才與谷歌建立合作。The work uses MedGemma, an open model hospitals can run inside their own systems, so records and images need not leave the estate. Singapore's AI partnership with Google only began in May.

顧問觀點

我們認為值得留意的是部署形態而非模型能力：資料留在院內，責任仍由醫院承擔。這種分工對亞太醫療機構是可複製的談判樣本。What merits attention is the deployment shape, not model capability: data stays in the hospital while the hospital carries the liability. That division is a replicable template for providers in Asia-Pacific.

對日常工作的影響: 醫療機構資訊主管要為院內部署的開源模型設計版本與更新流程，臨床部門訂明醫生覆核的觸發點，法務寫清跨境支援的資料邊界。Hospital IT leads should design versioning and update paths for on-premises open models, clinical leads should set the doctor review trigger, and legal should fix the data boundary for cross-border support.
如何改善: 本週就一個院內試點寫明模型版本、更新責任與人工覆核觸發點三行，交資訊與臨床雙方會簽。This week, for one in-house pilot, write the model version, update ownership and review trigger in three lines, signed off by IT and clinical leads.

來源騰訊新聞（2026-10-31）↗

日本、美國、澳洲與印度外長宣布出資逾八千五百萬美元，首次把印太可信人工智能落到預算Japan, the US, Australia and India pledge over 85 million dollars for trusted AI in the Indo-Pacific

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners法務合規Legal & compliance市場推廣Marketing

事件: 四國外長在亞特蘭大宣布出資逾八千五百萬美元（約一百三十四億日圓）支持印太可信人工智能，是四方安全對話機制首次把人工智能寫入出資科目。Foreign ministers of Japan, the United States, Australia and India, meeting in Atlanta, announced more than 85 million dollars, about 13.4 billion yen, to support development and use of trusted AI in the Indo-Pacific. It is the first time the Quad has attached a funding line to AI.
背後: 金額在算力投資中屬零頭，意義在機制：可信人工智能由聯合聲明變成可執行項目，評估方法與標準會隨資金一併輸出到受援市場。The sum is small against compute spending; the significance is institutional. Trusted AI moves from communique language to funded projects, and assessment methods travel with the money into recipient markets.

顧問觀點

我們認為企業要留意「可信」將被具體化為採購條款：評估框架隨資金落地後，供應商在這些市場投標須提交模型與數據治理證明，比任何自願指引有效。We expect "trusted" to be made concrete as procurement conditions. Once frameworks arrive with the funding, vendors bidding in these markets must submit model and data governance evidence, which binds harder than voluntary guidance.

對日常工作的影響: 法務與資訊主管要把預期評估要求前置到供應商問卷，市場部門在投標文件加一節說明模型來源與數據處理，採購前線開始索取證明。Legal and IT leads should pre-load likely assessment requirements into vendor questionnaires, marketing should add a section on provenance and data handling to bids, and procurement should start requesting evidence.
如何改善: 本週向兩個主要人工智能供應商索取一份數據治理與模型評估證明，建立存檔範本。This week, request a data governance and model assessment attestation from your two main AI vendors and set up a filing template.

來源機械姬安娜（引共同社）（2026-10-31）↗

本週可做的三件事

把「大規模部署人工智能前告知受影響員工」寫進下一次流程改動的立項表，並指定一名負責人。Add advance notice to affected staff, ahead of large-scale AI deployment, to the intake form for your next process change, and name one accountable owner.
在官網與各子網域統一更新 robots.txt，加入 __usage_ai-training 偏好行並保留原有爬蟲規則。Update robots.txt across the corporate site and all subdomains with a __usage_ai-training preference line, keeping existing per-crawler rules in place.
抽查三個月前做過的定價測試，確認輸入特徵裡沒有可反推收入、住址或健康狀況的代理變數。Sample any pricing test run in the past three months and confirm the input features contain no proxy that reveals income, address or health status.

常見問題

今日（2026-11-01）「AiX 環球 AI 情報」有哪十條重點？

今日十條為：1、國務院辦公廳首份人工智能就業文件出台，六方面十八項任務寫入失業緩衝期與應用前告知；2、OpenAI 把本輪估值目標由一點四萬億美元下調至一點二五萬億，募資額減至二百億；3、美國聯邦貿易委員會向九家企業發出六(b)調查令，追查算法如何決定價格與零工報酬；4、Chainalysis 承認遭「攻擊型代理技術」入侵，客戶姓名、電郵與職務標籤被讀取；5、OpenAI、谷歌、微軟與 Anthropic 同認一個 robots.txt 訓練退出令牌，生效時間各異；6、北京市出台智能體措施，「模型首方案」最高補貼三千萬元、目標算力十萬 P；7、Anthropic 調整 Claude 訂閱：Max 20x 更名 Ultra 並取消五小時限額，每週額度不變；8、韓國個人資訊保護委員會修訂安全措施標準，第四、五、六、八條於十月三十一日生效；9、谷歌在新加坡設研究中心並與兩大醫療集團合作，以開源模型在院內運行臨床助手；10、日本、美國、澳洲與印度外長宣布出資逾八千五百萬美元，首次把印太可信人工智能落到預算。每條均附本會顧問觀點、對日常工作的影響與一項可即時執行的建議。

本會對今日 AI 局勢的整體判斷是甚麼？

今日四條線指向同一件事：誰有權讓算法做什麼。國務院辦公廳首份回應人工智能就業效應的文件，把失業緩衝期、應用前告知寫成可執行義務；北京以三千萬元補貼模型首方案；美國聯邦貿易委員會要九家企業交代算法如何定價；Chainalysis 承認被攻擊型代理技術攻破。權限與紀錄已取代跑分。

這些情報的資料來源是甚麼？如何核實？

本欄每日由 AIX Society 編輯部檢索公開資料後撰稿，每條新聞均附原始來源連結（本期包括：中國經營報、中國新聞社、Reuters、中新經緯、Finks AI、The Hacker News、InYourGeek、Emergent），並將事實與本會觀點分列。讀者可按來源連結自行核實。

訂閱每週電子報

每週一封：AI 轉型的實戰觀察、工具實測與案例拆解。