一部安裝在公司伺服器上的代理,可以自己讀取收件箱、自己打開客戶檔案、自己把一份報價寫進系統,中間沒有一個鍵被人按過。2026 年 8 月,個人資料私隱專員公署發布《人工智能(AI):個人資料保障模範框架》的代理式人工智能篇《保障個人資料私隱》,文件第二頁回答了一個很多公司在採購時沒有問的問題:當智能體替公司做事,哪一方是《私隱條例》下的資料使用者[1]。An agent running on a company server can read an inbox, open a client file and write a quotation into a system without a key being pressed. In August 2026 the Office of the Privacy Commissioner for Personal Data published Artificial Intelligence: Model Personal Data Protection Framework and its agentic AI supplement. Its second page answers a question many companies never ask while buying: when an agent acts for the company, which party is the data user under the Personal Data (Privacy) Ordinance[1].
答案是公司。智能體沒有法律實體的地位,它不承擔責任,也不能作為資料使用者;透過它收集、持有、處理或使用個人資料的機構,仍然是資料使用者,六項保障資料原則照樣適用[1][4]。所以真正的決定不在付款那一刻,而在付款之前:代理讀得到哪幾個資料夾,寫得進哪幾個欄位,哪幾個動作要有人簽名,出事之後查得到哪幾行紀錄。The answer is the company. An agent has no legal personality: it does not carry responsibility and it cannot be a data user. The organisation that collects, holds, processes or uses personal data through it remains the data user, and the six data protection principles apply as they always did[1][4]. The decision therefore sits before payment, not at it: which folders the agent may read, which fields it may write, which actions need a signature, and which log lines survive an incident.
指引列出九項建議,附件 A 另有一份安全檢查清單,共 32 項,分五個階段。把 32 項按階段分開,上線之前佔 12 項,上線當日與之後佔 20 項,單是使用階段一段已有 10 項,是五個階段之中最多的一段[1]。九項建議之中,只有第九項提到培訓,其餘八項要用權限、日誌、合約與文件落實。這個分佈是這份指引自己的立場:多數檢查在上線之後才開始。The guidance sets out nine recommendations and Annex A adds a safety checklist of 32 items across five stages. Split by stage, 12 items fall before go-live and 20 fall on the day of deployment and after it; the in-use stage alone holds 10, more than any other stage[1]. Of the nine recommendations only the ninth mentions training; the other eight are delivered through permissions, logs, contracts and documents. That split is the guidance's own position: most of the checks begin after go-live.
摘要Summary
- 代理式 AI 在法律上沒有實體身分,透過它收集、持有、處理或使用個人資料的機構仍然是資料使用者,六項保障資料原則照樣適用[1][4]。Agentic AI has no legal personality, so the organisation that collects, holds, processes or uses personal data through it remains the data user and the six data protection principles still apply[1][4].
- 指引列出五類風險:存取權限過寬、系統漏洞、插件與技能帶來的惡意程式、資料被用於新目的,以及多重代理之間把錯誤資料層層傳遞[1]。The guidance names five risks: excessive access, system vulnerabilities, malicious plugins or skills, data reused for a new purpose, and false data passed between multiple agents[1].
- 九項建議之中,七項直接對應六項保障資料原則的其中一項;第八項要求持續評估風險;第九項要求分配責任、以合約約束供應商並提供培訓[1]。Seven of the nine recommendations map directly onto one of the six data protection principles; the eighth requires ongoing risk assessment; the ninth requires allocated responsibility, contract terms for suppliers and training[1].
- 附件 A 的檢查清單共 32 項,分評估、準備、部署、使用與停用五個階段;上線之前 12 項,上線當日與之後 20 項,使用階段一段已有 10 項[1]。The Annex A checklist holds 32 items across five stages, assessment, preparation, deployment, use and decommissioning: 12 before go-live, 20 on the day and after it, and 10 in the in-use stage alone[1].
- 批准向第三方送出個人資料、永久刪除個人資料與修改系統設定,三類操作都要有人在決策過程中保留最終控制權[1]。Approving the release of personal data to a third party, permanently deleting personal data and changing system settings each need a person to keep final control of the decision[1].
- 日誌要答得出五件事:做過甚麼、幾時做、用哪一個工具、碰過哪些資料、為甚麼這樣決定[1]。A log has to answer five questions: what was done, when, with which tool, on which data, and why the decision was made[1].
- 九項之中,一場培訓做得到的是第九項的一部分,加上第一至第五項的判斷;權限、日誌、合約與刪除時限要在課室以外完成[1]。A workshop can deliver part of the ninth recommendation and the judgement behind the first five; permissions, logs, contracts and deletion deadlines are finished outside the classroom[1].
代理式 AI 的事,法律上由誰負責?Who is responsible when an agent acts?
負責的是使用它的機構。指引寫明,智能體沒有法律實體的地位,透過智能體收集、持有、處理或使用個人資料的機構,作為資料使用者仍須負責遵守《私隱條例》下的規定,包括六項保障資料原則[1]。條例第 2(1) 條把資料使用者定義為獨自或聯同其他人控制個人資料的收集、持有、處理或使用的人;代理不在這個定義之內,因為它沒有可以承擔責任的身分[4]。The organisation using it. The guidance states that because an agent is not a legal entity, the organisation that collects, holds, processes or uses personal data through it remains the data user and must still comply with the Ordinance, including the six data protection principles[1]. Section 2(1) defines a data user as a person who controls the collection, holding, processing or use of the data, alone or jointly with others; an agent falls outside that definition because it has no standing to carry responsibility[4].
這句話把採購問題換成另一個問題。公司平常比較的是功能、價錢與整合,指引要求公司先答的是:這個代理可以做甚麼,不可以做甚麼,誰批准它做,做完之後留下甚麼。前一個問題在簽約前有答案,後一個問題要一直答到停用那一日。指引在 2026 年 8 月 25 日發布,同日的新聞稿同時交代了三宗個案:三名外籍家庭傭工把僱主的姓名、住址與電話號碼提供給財務機構申請貸款,僱主收到追數電話與逾期繳款通知書;專員認為三人違反保障資料第 3(1) 原則,向她們發出警告信[2]。That sentence turns a procurement question into a different one. Companies usually compare features, price and integration; the guidance asks first what this agent may do, what it may not do, who approves it, and what record survives the work. The first question can be answered before signing; the second has to be answered until the day the agent is switched off. The guidance was published on 25 August 2026, and the same day's press release carried three cases: three foreign domestic helpers gave their employers' names, addresses and telephone numbers to finance companies to apply for loans, the employers received collection calls and overdue notices, and the Commissioner found the three had breached data protection principle 3(1) and issued warning letters[2].
三宗個案與代理無關,但它們說的是同一條原則。保障資料第 3(1) 原則限制個人資料的使用,包括披露與轉移,只可以用於收集時擬使用的目的或直接有關的目的,除非資料當事人自願給予訂明同意[4]。一個代理把送貨用的客戶地址帶去推銷,或把查詢紀錄帶去訓練模型,踩到的就是這一條。用途一變,同意要重新取得,而同意不會因為操作由機器完成而消失。The three cases involve no agent, and they rest on the same principle. Data protection principle 3(1) limits the use of personal data, including disclosure and transfer, to the purpose for which it was collected or a directly related purpose, unless the data subject gives prescribed consent voluntarily[4]. An agent that carries a delivery address into a marketing list, or a query record into model training, hits that principle. Change the purpose and consent has to be obtained again, and consent does not disappear because a machine did the work.
公署在五個月之前已經就同一件事發過一次提醒。2026 年 3 月 16 日的新聞稿點名 OpenClaw 一類代理式工具,列出五點:授予最小權限、使用官方最新版本、把運行環境隔離並收窄網絡暴露面、審慎安裝插件與技能、持續評估風險並在重大影響的決定上採取人在環中[3]。八月那一份把同樣的風險寫成九項建議,再補上一份可以逐格打勾的清單。警告在前,指引在後,中間隔了五個月,而這五個月正是很多公司開始試用代理的時候。The Commission had already warned about the same shift five months earlier. Its statement of 16 March 2026 named OpenClaw and similar agentic tools and set out five points: grant the least privilege, use the official and current version, isolate the runtime and narrow the network exposure, install plugins and skills with care, and keep assessing risk while keeping a person in the loop on decisions with serious consequences[3]. The August document turns the same risks into nine recommendations and adds a checklist that can be ticked item by item. The warning came first and the guidance second, five months apart, and those five months are when many companies started their trials.
供應商在這一條線上的位置要分清楚。供應商只按公司的指示處理資料,它就是資料處理者,公司用合約約束它;供應商一旦為自己的目的使用同一批資料,例如用來訓練模型或改善自己的產品,它在那一部分就同時是資料使用者,要自己遵守條例[1][4]。公司要問的問題因此有兩層:這批資料由誰控制,供應商會不會把它用在自己的目的上。第二層的答案通常寫在服務條款的中段,而不在報價單上。The supplier's place on that line needs care. A supplier that processes data only on the company's instructions is a data processor, and the company binds it by contract; a supplier that uses the same data for its own purposes, such as training a model or improving its product, is also a data user for that part and has to comply in its own right[1][4]. The company therefore asks two questions: who controls this data, and will the supplier use it for a purpose of its own. The answer to the second usually sits in the middle of the terms of service, not on the quotation.
五類風險在公司裡的樣子What the five risks look like inside a company
指引把風險分成五類,每一類都可以在一個普通辦公室裡找到對應的畫面。表一左邊是文件用的名稱,中間是它在公司裡的樣子,右邊是九項建議之中回應它的那一項。五類風險全部落在權限、版本與紀錄三件事上,換一個更貴的工具不會令它們消失。The guidance groups the risks into five classes, and each has a picture inside an ordinary office. Table 1 gives the document's name on the left, the way it looks in a company in the middle, and the recommendation that answers it on the right. All five land on permissions, versions and records, and a more expensive tool does not remove one of them.
| 風險Risk | 在公司裡會見到甚麼How it appears | 回應它的建議Answered by |
|---|---|---|
| 廣泛的存取權限Excessive access | 代理讀得到整個共用資料夾,包括人事檔案與客戶合約;它亦有刪除檔案的權限,一次誤解指令就可以移除一批紀錄。The agent can read the whole shared drive, including HR files and client contracts, and it may delete files, so one misread instruction can remove a batch of records. | 第一項的分隔措施與第六項(d)的最小權限The segregation in the first recommendation and the least privilege in the sixth (d) |
| 系統漏洞System vulnerabilities | 同事用的是舊版本,或由第三方網站下載的安裝檔,漏洞未被修補。Staff run an old version, or an installer downloaded from a third-party site, and the vulnerability stays unpatched. | 第六項(a)與(b)的官方版本與環境隔離The official version and isolation in the sixth (a) and (b) |
| 插件或技能的漏洞Plugin and skill vulnerabilities | 有人裝了一個來源不明的技能,程式碼其後接管了帳戶,甚至整部電腦。Someone installs a skill from an unknown source, and the code later takes over the account or the whole machine. | 第六項(c)的來源審核The source check in the sixth (c) |
| 資料用途改變A changed purpose | 對話紀錄與長期記憶被用來訓練模型,超出收集資料時講明的目的。Conversation records and long-term memory are used to train a model, beyond the purpose stated when the data was collected. | 第五項的新目的限制,並在第二項的聲明中交代The limit on new purposes in the fifth, disclosed under the second |
| 多重代理系統的風險Multiple agents | 一個代理的錯誤資料傳給另一個代理,最後寫進客戶紀錄,而當事人難以行使查閱與改正的權利。One agent's wrong data passes to another and ends up in the client record, and the person concerned struggles to exercise access and correction rights. | 第三項的準確性與第八項的人在環中The accuracy in the third and the human oversight in the eighth |
右邊一欄有一個共同點:五個回應全部落在權限與紀錄之上。一個代理如果只讀得到三份檔案、只寫得進一個欄位、每一個改動都留一行紀錄,五類風險之中有四類會即時收窄;剩下的系統漏洞交給版本更新與隔離。反過來說,一個讀得到整個資料夾、又有刪除權限的代理,就算裝了最新的防護,它做錯事的代價仍然由公司支付。The right-hand column shares one property: all five answers sit in permissions and records. An agent that can read three files, write one field and leaves a line for every change narrows four of the five risks at once, and version updates and isolation handle the rest. An agent that reads a whole drive and may delete files still puts the cost of its mistakes on the company, whatever protection ships with it.
多重代理那一類風險值得一個具體畫面。假設一間公司用三個代理:一個讀取客戶電郵並摘錄要求,一個把要求寫進訂單系統,一個在缺貨時通知客戶。第一個代理把「14 日內交貨」聽成「4 日內交貨」,錯誤資料傳到第二個代理,第二個代理照寫,第三個代理照通知,客戶收到一個公司根本做不到的承諾。三個代理都沒有壞,每一個都按指示完成自己的工作,錯誤在傳遞之間被放大,而唯一能攔住它的是第二個代理寫入之前的一道核對[1]。The multiple-agent risk deserves a picture. A company runs three agents: one reads client email and extracts the requirement, one writes the requirement into the order system, and one tells the client when stock is short. The first agent mishears fourteen days as four, the wrong figure passes to the second, the second writes it, the third repeats it, and the client receives a promise the company cannot keep. None of the three agents is broken and each did its own job; the error grew in the handover, and the only thing that could have stopped it is a check before the second agent writes[1].
九項建議與六項保障資料原則The nine recommendations and the six principles
九項建議按文件本身的次序排,表二把它們逐項列出,右邊兩欄是它對應的保障資料原則,與公司在這一項之下要交出的一份文件。第四欄是整張表的用途:公司不需要記住九個標題,只需要答得出九份文件在哪裡。The nine recommendations keep the document order. Table 2 lists them with the principle each one maps to and the document the company produces under it. The fourth column is the point of the whole table: a company does not need to memorise nine headings, it needs to know where nine documents live.
| 建議Recommendation | 對應的原則Principle | 要交出的一份文件The document it produces |
|---|---|---|
| 1避免過度或任意收集個人資料1Avoid excessive or arbitrary collection | 第 1(1) 原則DPP 1(1) | 一份存取範圍清單:哪個資料夾、哪個系統,可讀或可寫A scope list: which folder and which system, read or write |
| 2就代理處理個人資料保持透明度2Be transparent about the processing | 第 1(3) 及第 5 原則DPP 1(3) and 5 | 收集個人資料聲明與私隱政策各加一段One new paragraph in the collection statement and the privacy policy |
| 3確保代理處理的個人資料準確3Keep the data the agent handles accurate | 第 2 原則DPP 2 | 一份核實規則:哪幾類輸出一律由人手核對A verification rule: which outputs always get a human check |
| 4訂定適當的保留時限4Set a retention limit | 第 2(2) 原則DPP 2(2) | 對話紀錄、快取與長期記憶的刪除期限Deletion deadlines for conversations, cache and long-term memory |
| 5未經同意不得用於新目的5No new purpose without consent | 第 3 原則DPP 3 | 一句寫明代理只可為哪些目的處理個人資料One sentence stating the purposes the agent may serve |
| 6保障代理系統內的個人資料安全6Secure the data inside the agent | 第 4 原則DPP 4 | 權限表、日誌設定與高風險操作清單A permission table, the log settings and the high-risk list |
| 7確保查閱及改正資料的權利7Protect access and correction rights | 第 6 原則DPP 6 | 一條在一日內交出某個人資料的流程A route that produces one person's data within a day |
| 8持續進行風險評估8Keep assessing the risk | 文件未指定原則No principle named | 上線前的測試紀錄與一份高風險操作名單The pre-launch test record and the high-risk list |
| 9分配明確責任及提供培訓9Allocate responsibility and train staff | 文件未指定原則;合約要覆蓋第 2 及第 4 原則No principle named; the contract covers DPP 2 and 4 | 一個具名職位、三條合約條款、一份培訓紀錄A named role, three contract clauses and a training record |
九項按性質分開,是五項要寫的文件、一項要改的系統設定、兩項要走的流程與一項要指派的責任。九項之中只有第六項是技術設定,它把六個動作寫在同一項之下:用官方最新版本、隔離運行環境、審核插件與技能、授予最小權限、加安全護欄、開啟日誌[1]。其餘八項都可以由一個人在一張會議桌上完成,前提是他知道要寫哪一份文件。Sorted by kind, the nine are five documents to write, one set of system settings to change, two processes to run and one responsibility to assign. Only the sixth is a technical setting, and it gathers six actions: use the official current version, isolate the runtime, vet plugins and skills, grant the least privilege, add guardrails and switch on logging[1]. The other eight can be done by one person at one meeting table, provided that person knows which document to write.
第七項最容易被當成資訊科技部的問題,它其實是一條時限。代理的資料散在對話紀錄、快取與長期記憶三處,一個客戶要求查閱自己的資料時,公司要在一日之內找齊三處的紀錄,還要判斷哪一條是準確的版本。指引因此建議選用在設計階段已納入貫徹私隱設計與預設私隱的系統,並在採購時問供應商同一個問題[1]。一個答不出這個問題的代理,上線之後每一次查閱都會變成一次人手搜索。The seventh looks like an IT problem and is really a deadline. An agent's data sits in three places, the conversation record, the cache and long-term memory, so when a client asks for their data the company has to gather all three within a day and decide which version is accurate. The guidance therefore suggests choosing systems designed with privacy by design and privacy by default, and asking suppliers the same question at purchase[1]. An agent that cannot answer it turns every access request into a manual search after go-live.
第二項要求的透明度可以收成一段話,放進收集個人資料聲明:「本公司使用人工智能代理處理客戶查詢,過程中會讀取閣下的姓名、聯絡方法與查詢內容;代理不會把資料用於其他目的,紀錄的保存期限為 24 個月。」指引同時建議在私隱政策聲明中述明資料儲存或處理所在的地區,公司在使用境外供應商時尤其要寫清楚[1]。一段話寫得出來,第二項就交了;寫不出來,通常代表公司自己也未清楚代理讀過甚麼。The transparency in the second recommendation condenses into one paragraph for the collection statement: this company uses an AI agent to handle client enquiries, the agent reads the enquirer's name, contact details and the enquiry, it uses the data for no other purpose, and the record is kept for 24 months. The guidance also suggests stating where the data is stored or processed, which a company using an overseas supplier needs more than most[1]. If the paragraph can be written, the second recommendation is done; if it cannot, the company usually does not know what the agent has read.
檢查清單的五個階段The checklist's five stages
附件 A 把 32 項檢查分成五個階段,表三按次序列出五段的項數與各自最容易被跳過的一項,每一項前面有一個方格,右邊兩欄分別標示它適用於機構還是個人。32 項之中有 30 項同時適用於機構與個人,兩項屬於機構獨有,兩項都在準備階段:一項要機構在收集個人資料聲明與私隱政策聲明中交代代理的用途,另一項要機構建立管治架構並為人員提供培訓[1]。換句話說,同事在自己電腦上安裝同類工具,也逃不出這份清單的範圍。Annex A splits the 32 checks into five stages, and Table 3 sets out the item count of each with the one most often skipped, each check with a box to tick and two columns marking whether it applies to an organisation or an individual. Thirty of the 32 apply to both, and the two that do not are both in the preparation stage: one has the organisation disclose the agent's use in its collection statement and privacy policy, the other has it set up a governance structure and train staff[1]. A colleague installing a similar tool on a personal machine is still inside the checklist's reach.
| 階段Stage | 項數Items | 公司要交出的一份文件What the company produces | 最常被跳過的一項The item most often skipped |
|---|---|---|---|
| 評估Assessment | 55 | 一份寫明用途與風險的評估紀錄A record of the purpose and the risk | 不使用存在顯著不合理安全風險的代理Do not use an agent with an unreasonable risk |
| 準備Preparation | 77 | 收集個人資料聲明與私隱政策的新段落The new paragraphs in the two statements | 從官方渠道取得安裝檔案Obtain the installer from the official channel |
| 部署Deployment | 77 | 日誌設定與最小權限的紀錄The log settings and the least-privilege record | 提前定義高風險操作Define the high-risk operations in advance |
| 使用Use | 1010 | 每月覆核紀錄與長期記憶的清理紀錄The monthly review and the memory clean-up record | 定期手動查閱長期記憶的內容Read the long-term memory by hand, regularly |
| 停用Decommission | 33 | 一份停用時的資料匯出與清理紀錄The export and clean-up record at shutdown | 解除安裝前匯出需要保存的數據Export the data before uninstalling |
五個階段的項數是 5、7、7、10、3。上線之前只有 12 項,上線當日與之後有 20 項:評估與準備合共 12,部署、使用與停用合共 20。使用階段一段有 10 項,佔全部 32 項的三成,是五個階段之中最多的一段[1]。這個分佈與多數公司的預算分配相反:預算放在採購,覆核留給「之後再算」,而指引把超過六成的檢查放在上線之後。The five stages hold 5, 7, 7, 10 and 3 items. Twelve items fall before go-live and 20 on the day of deployment and after it: assessment and preparation make 12, deployment, use and decommissioning make 20. The in-use stage alone holds 10, 30 per cent of the 32 and the largest share of any stage[1]. That distribution runs against most budgets, which spend on procurement and leave the review for later: more than sixty per cent of the checks sit after go-live.
使用階段的十項之中,有五項寫明要定期做:覆查安裝與配置、檢查對外介面、備份重要資料、手動查閱長期記憶,以及清理不再使用的技能與對話紀錄。五項之間有一個共同要求,就是做過之後要留下日期與簽名,否則下一次覆核沒有起點。指引把使用階段放在部署之後,正是因為代理的記憶會纍積:第一日它只知道三份檔案,第三個月它記得二十個客戶的對話。Five of the ten in-use items are explicitly periodic: recheck the installation and configuration, test the external interfaces, back up the important data, read the long-term memory by hand, and clear out skills and conversation records no longer in use. All five require a date and a signature afterwards, or the next review has no starting point. The guidance puts the in-use stage after deployment precisely because an agent's memory accumulates: on day one it knows three files, and by the third month it remembers the conversations of twenty clients.
停用階段只有三項,卻是最少公司預先安排的一段。三項的次序是先停止使用,在解除安裝之前匯出或轉存需要保存的數據,最後清理運行環境[1]。一個代理用了兩年之後,長期記憶裡可能有幾百段對話,當中夾雜客戶的個人資料;沒有匯出與清理的紀錄,公司在停用的時候就把一個資料庫留在硬碟上。指引把停用寫成一個獨立階段,等於要求公司在第一天就想好最後一天。The decommissioning stage holds only three items and is the one companies plan least. The order is to stop using the agent, export or transfer the data that has to be kept before uninstalling, and clean the runtime[1]. After two years an agent's long-term memory may hold hundreds of conversations with client personal data inside them, and without an export and clean-up record the company leaves a database on a disk when it switches the agent off. The guidance makes decommissioning a stage of its own, which asks the company to think about the last day on the first one.
代理可以自己刪除客戶資料嗎?May the agent delete client data on its own?
可以,但公司要事先寫明哪幾種刪除要有人批准。指引在部署階段要求機構提前定義高風險操作,第八項則要求在代理的決定可能對個人產生重大影響時保留最終控制權,文件舉了三個例子:批准向第三方發送個人資料、永久刪除個人資料,以及修改系統配置[1]。三個例子有一個共通點,做完之後收不回來,或者資料已經離開公司。表四把三個例子連同插件安裝寫成一張可以貼在會議室牆上的規則表,第四行來自第六項(c)對插件來源的要求。Yes, provided the company has written down in advance which deletions need approval. The deployment stage asks organisations to define high-risk operations in advance, and the eighth recommendation asks them to keep final control where an agent's decision may seriously affect a person; the document gives three examples: approving the release of personal data to a third party, permanently deleting personal data, and changing system settings[1]. All three share one property: after the act, the data cannot be recalled or it has already left the company. Table 4 turns the three, with plugin installation, into a rule sheet that can go on a meeting-room wall; the fourth row comes from the sixth recommendation's requirement to vet plugin sources.
| 操作Operation | 為甚麼要人手Why a person signs | 由誰簽Who signs | 留下甚麼What is left behind |
|---|---|---|---|
| 向第三方送出個人資料Releasing data to a third party | 資料一離開公司就收不回,而收件人未必受同一份合約約束。Data cannot be recalled once it leaves, and the recipient may not be bound by the same contract. | 業務主管與資訊科技主管The business lead and the IT lead | 一則帶日期的批准、收件人,與送出的欄位清單A dated approval, the recipient and the list of fields sent |
| 永久刪除個人資料Permanently deleting data | 刪除不可逆轉,亦可能與保留政策或查閱要求相撞。Deletion cannot be undone, and it may collide with a retention rule or an access request. | 負責個人資料的職位The role responsible for personal data | 刪除清單、批准人與時間The deletion list, the approver and the time |
| 修改系統配置Changing system settings | 權限一改,之前所有紀錄的意思都跟着改變。A change of permission changes what every earlier record means. | 資訊科技主管The IT lead | 改動前後的設定與改動原因The setting before and after, and the reason |
| 安裝插件或技能Installing a plugin or skill | 程式碼可以在代理的權限之內做任何事,包括接管帳戶。Code can do anything inside the agent's permissions, including taking over the account. | 資訊科技主管The IT lead | 來源、版本與安全檢查結果The source, the version and the security check |
「由誰簽」那一欄要寫職位,不寫名字。職位在換人的時候不會失效,而代理的權限不會因為換人而自動收窄。表四的四類操作可以壓縮成一條規則:代理可以準備,人可以送出。這條規則在一間四十人的公司行得通,在一間四百人的公司亦行得通,分別只在於第二個人是誰。The signature column names a role, not a person. A role survives a change of staff, and an agent's permissions do not narrow themselves when the staff change. The four operations in Table 4 compress into one rule: the agent prepares and a person sends. The rule works in a forty-person company and in a four-hundred-person one; only the name of the second person changes.
日誌要記錄的五個欄位The five fields a log has to keep
第六項(f)要求建立機制,確保代理的可追溯性與可審核性,並舉了五樣要記下的事:智能體做過的操作、操作的時間、所使用的工具、存取過的資料,以及決策理由[1]。五樣之中,公司最容易漏掉的是最後一樣。操作與時間由系統自動寫入,決策理由只有人寫得出來,而事故之後最常要回答的正是這一條:當時為甚麼決定這樣做。Recommendation six (f) asks for a mechanism that makes an agent traceable and auditable, and names five fields to record: the operation performed, its time, the tool used, the data accessed and the reason for the decision[1]. The fifth is the one companies drop first. The operation and the time are written by the system, only a person can write the reason, and the question an incident asks first is exactly that one: why was this the decision at the time.
示範紀錄:一行日誌,五個欄位A sample record: one log line, five fields
| 欄位Field | 一行填好的例子How one line fills it |
|---|---|
| 操作Operation | 讀取三份報價 PDF,把金額與交貨期寫入報價系統的草稿Read three quotation PDFs and wrote the amounts and delivery dates into a draft in the quotation system |
| 時間Time | 2026-11-19 14:32,由系統寫入19 November 2026, 14:32, written by the system |
| 工具Tool | 報價助手 v2,官方版本,未安裝任何第三方技能Quotation assistant v2, the official build, with no third-party skill installed |
| 存取的資料Data accessed | 客戶聯絡人姓名、公司地址、報價金額;未觸及身份證號碼與銀行資料Client contact name, company address and the quoted amount; no identity number and no bank details |
| 決策理由Reason | 客戶要求 15% 折扣,超出代理的授權,草稿留給陳小姐決定The client asked for 15 per cent off, beyond the agent's authority, so the draft was left for Ms Chan to decide |
這一行日誌答得出五條問題,而它只花了寫的人二十秒。日誌的作用有三個:事故之後追查範圍、查閱請求來的時候找齊紀錄、覆核的時候看得出代理的權限有沒有被人改過。三個作用都要靠同一件事,就是日誌由系統寫,理由由人寫,兩者在同一行。分開存放的日誌,在第一次查閱請求時就會露底。This line answers five questions and took twenty seconds to write. A log serves three purposes: scoping an incident afterwards, gathering the records when an access request arrives, and showing at review whether the agent's permissions were changed. All three depend on one property: the system writes the log and a person writes the reason, on the same line. A log kept in two places fails at the first access request.
簽了合約,責任會轉給供應商嗎?Does the contract move the duty to the supplier?
不會。合約可以令供應商承擔更多,公司在條例下的身分不會因為一份合約改變。指引第九項要求機構在委聘外部服務供應商時,採取合約規範方法或其他方法,確保資料處理者遵守第 2 原則的保留規定與第 4 原則的保安規定[1]。這一句話的重量在「確保」兩個字:條款寫得出來,公司還要證明它執行得到。No. A contract can put more on the supplier, and the company's standing under the Ordinance does not change with it. The ninth recommendation asks an organisation engaging an external supplier to use contractual means or other means to ensure the data processor complies with the retention rule in principle 2 and the security rule in principle 4[1]. The weight sits in the word ensure: a clause can be written, and the company still has to show it is enforced.
三條條款覆蓋第九項要求的大部分內容,每一條都可以在採購會議上讀出來,對方聽得明,公司亦驗得到。Three clauses cover most of what the ninth recommendation asks, and each can be read aloud in a procurement meeting so the supplier understands it and the company can check it.
- 第一條:刪除與交回。代理停用或合約結束之後,供應商在指定日數內刪除或交回公司資料,並以書面確認;分包商名單同時更新。Clause one: deletion and return. After the agent is switched off or the contract ends, the supplier deletes or returns the company's data within a stated number of days and confirms it in writing, and the list of subcontractors is updated at the same time.
- 第二條:事故通報。供應商發現外洩或未授權存取之後,在指定小時內通知公司,通知要說清楚哪些資料、涉及多少人、已經採取了甚麼措施。Clause two: incident notice. Once the supplier discovers a leak or unauthorised access, it tells the company within a stated number of hours, naming the data involved, the number of people affected and the steps already taken.
- 第三條:紀錄與審核。公司有權取得與代理操作有關的日誌,並在事前通知之下審核系統與分包商,審核結果以書面交付。Clause three: records and audit. The company may obtain the logs of the agent's operations and audit the system and its subcontractors on notice, with the findings delivered in writing.
第二條最能夠分辨一個供應商。願意寫明通報時限的供應商,通常已經有事故流程;不願意寫的,往往連內部由誰負責通報都答不出來。公司在這一條上讓步,等於把最需要速度的那幾個鐘交給一個沒有承諾的人。The second clause sorts suppliers faster than any demonstration. A supplier willing to put an hour count on its notice has an incident process; one that will not usually cannot say which post sends the notice. A company that concedes this clause hands the hours that matter most to someone who has promised nothing.
三條條款寫得再好,都改變不到一件事:資料仍然由公司控制,責任仍然由公司承擔。合約的功能是把供應商拉進同一個流程,讓公司在事故之後有追討的依據,亦讓供應商在事前知道要做甚麼。把合約當成責任轉移的工具,是最常見的誤解,而誤解的代價通常在上線之後第三個月出現。However well the three clauses are drafted, one thing does not change: the company still controls the data and still carries the duty. The contract's job is to pull the supplier into the same process, giving the company a basis to recover after an incident and telling the supplier what to do before one. Treating the contract as a transfer of responsibility is the commonest misreading in this whole area, and the cost of it usually arrives in the third month after go-live.
一份填好的代理登記表A filled agent register
前面幾張表的內容放在同一張紙上,就是公司要維護的那一頁。下面兩欄是一間四十人貿易公司的示範,一間賣家居用品、客戶包括本地零售商與少量海外買家的公司:第一個代理處理共用收件箱的查詢,第二個代理讀取報價文件並填寫報價表。兩欄都寫成可以照抄的長度,公司換上自己的名稱與日期就可以用。Put the contents of the earlier tables on one sheet of paper and you have the page a company maintains. The two columns below are a sample for a forty-person trading company selling household goods to local retailers and a few overseas buyers: the first agent handles enquiries in a shared inbox, the second reads quotation documents and fills the quotation form. Both are written at a length that can be copied, and a company can use them by replacing the names and dates.
示範登記表:兩個代理,六個欄位A sample register: two agents, six fields
| 欄位Field | 收件箱助手 v1Inbox assistant v1 | 報價助手 v2Quotation assistant v2 |
|---|---|---|
| 用途Purpose | 把客戶查詢分類,並起草回覆的初稿Sort client enquiries and draft a first reply | 讀取三份報價文件,填寫報價表的草稿Read three quotation documents and fill a draft quotation |
| 觸及的個人資料Personal data touched | 寄件人姓名、電郵、電話,以及查詢內容Sender name, email, telephone and the enquiry itself | 客戶聯絡人、公司地址、報價金額Client contact, company address and the quoted amount |
| 權限Permissions | 讀取共用收件箱;只可寫入客戶關係系統的備註欄Read the shared inbox; write only to the notes field in the CRM | 讀取專案資料夾;寫入報價系統的草稿區,不可以送出Read the project folder; write to the draft area of the quotation system, never send |
| 高風險操作High-risk operations | 向第三方送出資料、刪除郵件Releasing data to a third party, deleting mail | 折扣超過 10%、修改報價系統的設定A discount above 10 per cent, changing quotation settings |
| 批准人Who approves | 營運總監Operations director | 業務主管Sales lead |
| 覆核日Next review | 2027-02-28,每季一次28 February 2027, then quarterly | 2027-02-28,每季一次28 February 2027, then quarterly |
六個欄位填得滿,第一項的存取範圍、第五項的目的限制、第六項的權限與日誌、第八項的高風險操作,與第九項的具名責任就同時有紀錄。登記表的好處是它會過期:覆核日一欄寫在紙上,時間一到就有人要簽名,而簽名的人在簽之前會問一句「這個代理還在做同一件事嗎」。這一問,就是第三項要求的準確性檢查在真實場景裡的樣子。Six fields filled give the scope in the first recommendation, the purpose limit in the fifth, the permissions and logging in the sixth, the high-risk list in the eighth and the named responsibility in the ninth, all in one record. The register's advantage is that it expires: the review date is on paper, and when it arrives someone has to sign, and before signing that person asks whether the agent still does the same job. That question is what the accuracy check in the third recommendation looks like in a real office.
登記表亦可以當成採購清單用。公司要買第二個代理時,把它的名稱、用途與權限填在同一張紙上,如果有一欄填不出來,那個代理未準備好上線。填不出「高風險操作」的公司,通常亦未想清楚代理的授權界線;填不出批准人的公司,等於沒有職位為它負責。The register doubles as a purchase list. When a company buys a second agent, it fills in the name, the purpose and the permissions on the same sheet, and a field that cannot be filled says the agent is not ready. A company that cannot fill the high-risk row has not settled the agent's authority; a company that cannot name an approver has no post answering for the agent.
培訓在九項建議裡佔多少How much of the nine a training day covers
第九項是九項之中唯一提到培訓的一項,而它同時要求三件事:建立具備足夠資源、專業知識及決策權的內部管治架構;在委聘外部供應商時以合約規範;為所有相關人員提供足夠的培訓[1]。培訓是這一項的三分之一。2025 年 3 月的《僱員使用生成式AI的指引清單》把培訓放在四項支援措施之中,與定期向同事傳達政策、委派支援隊伍、建立反饋機制並列,清單本身處理的是政策要涵蓋哪五項內容[6]。兩份文件的位置一致:培訓令政策生效,它不代替政策。The ninth is the only one of the nine that mentions training, and it asks for three duties at once: an internal governance structure with enough resources, expertise and decision-making authority; contractual controls when an external supplier is engaged; and adequate training for everyone involved[1]. Training is a third of that recommendation. The March 2025 checklist for employees using generative AI puts training among four supporting measures, beside communicating the policy regularly, appointing a support team and building a feedback channel, while the checklist itself sets out the five headings the policy covers[6]. Both documents place it the same way: training makes a policy work, and it does not stand in for one.
一場設計得好的培訓做得到三件事,三件都可以在一日之內看到。A well-designed training day delivers three outcomes, and all three are visible within a day.
- 同事認得哪幾類資料不可以輸入,也認得事故之後找誰,於是第一項與第六項的規則在人的層面有了執行者。Staff recognise which kinds of data may not be entered and who to call after an incident, so the rules in the first and sixth recommendations have someone to execute them.
- 批准的人在邊界個案上有判斷依據,例如客戶要求把資料用在新用途,或代理要求一次例外,於是第五項與第八項的判斷有了人選。Approvers have a basis for judgement on borderline cases, such as a client asking for data to be used for a new purpose or an agent asking for an exception, so the judgement in the fifth and eighth recommendations has someone to make it.
- 公司留下一份出席與測驗紀錄,這一頁紙就是第九項所講的培訓的證據。The company keeps a record of attendance and a short test, and that sheet is the evidence of training the ninth recommendation asks for.
培訓做不到的四件事同樣清楚:它改不到一個共用的權限設定,開不到一個日誌欄位,寫不進一份供應商合約,亦訂不出一個刪除期限。四件事都要在課室以外完成,而且每一件都比一場培訓便宜。一間公司如果只買一場培訓,它做到的只是第九項的三分之一,其餘八項原封不動。The four tasks training cannot do are just as clear: it cannot change a shared permission, switch on a log field, write itself into a supplier contract or set a deletion deadline. All four happen outside the classroom, and each costs less than a training day. It reaches a third of the ninth recommendation and leaves the other eight untouched.
這幾句對我們不利,我們仍然寫出來,因為它決定了我們怎樣接一個項目。我們的企業培訓會先問三個問題:代理的權限表在哪裡、日誌開了沒有、合約由誰簽。三個問題有答案,培訓才有規則可以依附;三個問題沒有答案,我們會先做那三件事,再排課。這個次序有時令一單生意延後一個月,它亦令公司在第二季的時候不需要重做一次。These lines work against us and are here anyway, because that fact decides how we take on a project. Our corporate training starts with three questions: where is the agent's permission table, is the log switched on, and who signs the contract. With answers, the training has a rule to attach to; without them, we finish those three first and schedule the class afterwards. That order sometimes delays a sale by a month, and it saves the company from doing the work twice in the second quarter.
公署在 2025 年 3 月的新聞稿表示,會舉辦講座介紹該份僱員指引與 2024 年的模範框架,並繼續為機構舉辦內部培訓講座[6]。監管機構自己做培訓,說明培訓在它的框架裡屬於推廣與支援,與合規文件分開處理。我們按同一個分法設計課程:九十分鐘講權限、日誌與合約三件事怎樣互相支撐;半日加一小時實作,讓參加者用自己公司的流程填一次登記表;全日才加入代理的測試與事故演練。三種長度對應三種未完成的工作,報名之前應該先知道自己缺哪一種;四項服務的範圍與形式寫在企業 AI 方案那一頁。In its March 2025 release the Commission said it would hold talks on the employee checklist and the 2024 framework and keep running in-house seminars for organisations[6]. A regulator that teaches treats training as outreach and support, handled apart from the compliance documents. We design to the same split: ninety minutes on how permissions, logs and contracts hold each other up; half a day with an hour of practice in which participants fill a register using their own process; a full day before testing an agent and rehearsing an incident. The three lengths answer three different unfinished tasks, and a buyer should know which one is missing before booking; the scope and format of the four services sit on the AI for business page.
這份指引沒有處理的地方What the guidance leaves out
指引處理個人資料,其餘的風險留給其他文件。偏見只在第八項出現一次,作為採取人在環中的理由[1];歧視、知識產權、員工行為與代理產出的內容由誰負責,都不在它的範圍。它提供一份檢查清單,沒有一份可以直接發給同事的政策;政策要由公司自己寫,寫法在另一篇拆開。它亦在最後一頁寫明,文件只作一般參考,不構成法律意見[1]。The guidance deals with personal data and leaves other risks to other documents. Bias appears once, in the eighth recommendation, as the reason for keeping a person in the loop[1]; discrimination, intellectual property, staff conduct and who answers for what an agent produces all sit outside the document. It provides a checklist, not a policy that can be sent to colleagues; the company writes that policy itself, and the governance article sets it out part by part. The final page states that the document is general reference and not legal advice[1].
指引放在兩份文件之後,三份一層一層疊住。2024 年 6 月的《人工智能(AI):個人資料保障模範框架》處理 AI 策略與管治、風險評估與人為監督、模型的定製與系統的實施管理,以及與持份者的溝通四個範疇,是採購與使用任何 AI 系統的底層[5]。2025 年 3 月的僱員清單處理同事手上的工具,五項內容包括獲准使用的範圍、個人資料的保障、合法與合乎道德的使用、數據安全,與違反政策的後果[6]。2026 年 8 月的代理篇處理一個新的情況:工具開始自己做事。三份都沒有被取代,代理篇更明言模範框架的原則一般仍然適用[1]。The guidance sits behind two earlier documents and the three stack. The June 2024 model framework covers four areas, AI strategy and governance, risk assessment and human oversight, model customisation and system implementation and management, and communication with stakeholders, and it remains the base for buying and using any AI system[5]. The March 2025 employee checklist covers the tools in staff hands, with five headings: the approved scope, protection of personal data, lawful and ethical use, data security, and the consequences of breach[6]. The August 2026 agent guidance covers a new situation, a tool that acts on its own. None of the three replaces another, and the agent guidance says the framework's principles generally still apply[1].
剩下的一個問題由公司自己回答:代理的權限應該開到幾大。指引要求最小權限與清晰的分隔,沒有給出一個數字,因為數字取決於流程。這個問題與部門之間的分工有關,與公司現時在哪一個採用階段亦有關,兩者分別在五個層級那一篇與培訓之後那一篇拆開。代理上線之後停不下來的公司,多數在權限那一步開得太大,而這一件事在採購會議上只值一句問題。One question is left for the company: how far the agent's permissions should reach. The guidance asks for the least privilege and clear segregation and gives no number, because the number depends on the process. That question touches how departments divide the work and where the company sits in its adoption, which the article on the five levels and the one on what happens after training take apart. Companies whose agents will not stop after go-live usually opened permissions too wide, and that is a single question in a procurement meeting.
我們的位置亦要寫清楚。我們做企業培訓與到場導入,不做法律意見,亦不代公司與監管機構往來;需要法律意見的時候,公司應該找律師。我們的限制在另一邊:一場培訓之後,公司的風險不會因此降低,除非同一天有人改了一個設定、寫了一條條款、開了一個日誌。這句話我們在課堂上講得最多,也是客戶最常跳過的一句。學校讀者問同一個問題時,答案在學校 AI 使用指引那一篇,因為學校的資料當事人是學生與家長,界線由另一組規則劃定。Our position belongs here too. We run corporate training and on-site implementation, we do not give legal advice and we do not deal with the regulator on a company's behalf; where legal advice is needed, a lawyer should give it. Our limit runs the other way: after a training day the company's risk has not fallen unless, on the same day, someone changed a setting, wrote a clause and switched on a log. That is the sentence we say most often in a classroom, and the one clients skip most often. A school reader asking the same question finds the answer in the article on a school's AI use policy, because there the data subjects are students and parents and a different set of rules draws the line.
常見問題Questions
代理是外國公司的產品,條例仍然適用嗎?The agent is a foreign product. Does the Ordinance still apply?
同事用個人帳戶的 AI 工具,公司要負責嗎?Staff use AI tools on personal accounts. Is the company responsible?
公司只有五個人,32 項都要做嗎?The company has five people. Must all 32 items be done?
日誌由供應商保存,公司拿不到怎麼辦?The supplier keeps the log. What if the company cannot get it?
已經用了代理半年,現在補做可以嗎?The agent has been running for six months. Can the work be done now?
這份指引與 2024 年的模範框架有甚麼關係?How does the guidance relate to the 2024 framework?
公司要成立 AI 委員會嗎?Does the company need an AI committee?
下一步Where to go next
相關文章Related articles
資料來源Sources
- 個人資料私隱專員公署《人工智能(AI):個人資料保障模範框架》的代理式人工智能篇《保障個人資料私隱》,2026 年 8 月,共 8 頁,2026-11-19 核對。文中引用:第 2 頁引言與智能體沒有法律實體地位那一段,以及資料使用者的定義;第 3 頁的五類風險;第 4 至 6 頁的九項建議與各項所引的保障資料原則;第 6 頁關於個人用戶的第 1、3、4、5 及 6 項;第 7 頁附件 A 的 32 項檢查、五個階段與機構及個人兩欄;第 8 頁的免責聲明。Office of the Privacy Commissioner for Personal Data, Artificial Intelligence: Model Personal Data Protection Framework, the agentic AI supplement, August 2026, an eight-page document checked on 19 November 2026. Used here: the introduction and the sentence that an agent is not a legal entity on page 2, with the definition of a data user; the five risks on page 3; the nine recommendations and the principle each cites on pages 4 to 6; page 6 on recommendations 1, 3, 4, 5 and 6 applying to individual users; the 32 checks, five stages and the organisation and individual columns of Annex A on page 7; and the disclaimer on page 8. pcpd.org.hk
- 個人資料私隱專員公署新聞稿「私隱專員公署跟進三宗外傭不當使用僱主個人資料申請貸款的個案 並發布《代理式人工智能篇:保障個人資料私隱》指引」,2026 年 8 月 25 日,2026-11-19 讀取。文中引用:指引的發布日期、數字政策辦公室與香港應用科技研究院擔任支持機構、九項建議的新聞稿摘要、附件 A 的五個階段,以及三宗個案的經過與保障資料第 3(1) 原則的裁斷。Office of the Privacy Commissioner for Personal Data, press release on three cases of foreign domestic helpers misusing employers' personal data to apply for loans, issued together with the agentic AI guidance, 25 August 2026, read on 19 November 2026. Used here: the publication date, the Digital Policy Office and the Hong Kong Applied Science and Technology Research Institute as supporting organisations, the release's summary of the nine recommendations, the five stages of Annex A, and the facts of the three cases with the finding under data protection principle 3(1). pcpd.org.hk
- 個人資料私隱公署新聞稿「私隱專員公署關注OpenClaw及代理式AI的私隱風險 提醒機構及市民安全使用AI」,2026 年 3 月 16 日,2026-11-19 讀取。文中引用:代理式工具的用途描述,以及五點建議:授予最小權限、使用官方最新版本、採取足夠的系統與資料安全措施、審慎安裝插件或技能、持續評估風險並採取人在環中。Office of the Privacy Commissioner for Personal Data, press release on the privacy risks of OpenClaw and agentic AI, 16 March 2026, read on 19 November 2026. Used here: the description of what an agentic tool does, and the five points, which are least privilege, the official current version, adequate system and data security, careful installation of plugins or skills, and ongoing risk assessment with a person in the loop. pcpd.org.hk
- 《個人資料(私隱)條例》(第 486 章),2026-11-19 讀取。文中引用:第 2(1) 條「資料使用者」的定義,附表 1 的六項保障資料原則,以及保障資料第 3(1) 原則關於新用途與訂明同意的規定。Personal Data (Privacy) Ordinance, Chapter 486, read on 19 November 2026. Used here: the definition of a data user in section 2(1), the six data protection principles in Schedule 1, and the rule in principle 3(1) on new purposes and prescribed consent. elegislation.gov.hk
- 個人資料私隱專員公署《人工智能(AI):個人資料保障模範框架》,2024 年 6 月,2026-11-19 核對。文中引用:四個範疇,即 AI 策略及管治架構、風險評估及人為監督、實行 AI 模型的定製與 AI 系統的實施及管理,以及促進與持份者的溝通及交流。Office of the Privacy Commissioner for Personal Data, Artificial Intelligence: Model Personal Data Protection Framework, June 2024, checked on 19 November 2026. Used here: the four areas, which are AI strategy and governance, risk assessment and human oversight, customisation of AI models and implementation and management of AI systems, and communication and engagement with stakeholders. pcpd.org.hk
- 個人資料私隱專員公署《僱員使用生成式AI的指引清單》,2025 年 3 月;同日新聞稿「私隱專員公署發表(一)《僱員使用生成式AI的指引清單》及(二)俊思管理有限公司資料外洩事故的調查結果」,2025 年 3 月 31 日,2026-11-19 讀取。文中引用:政策要涵蓋的五項內容(獲准使用的範圍、個人資料私隱的保障、合法及合乎道德的使用及預防偏見、數據安全、違反政策的後果),以及四項支援措施,即定期傳達政策、提供培訓及資源、委派支援隊伍、建立反饋機制。Office of the Privacy Commissioner for Personal Data, checklist for employees using generative AI, March 2025, with the press release of 31 March 2025 that issued it alongside the investigation report into a data leak at Imaginex Management, read on 19 November 2026. Used here: the five headings the policy covers, which are the approved scope, the protection of personal data, lawful and ethical use with bias prevention, data security and the consequences of breach, and the four supporting measures, which are communicating the policy, providing training and resources, appointing a support team and building a feedback channel. pcpd.org.hk
首次發表:2026-11-19,最後更新:2026-11-19First published 19 November 2026; last updated 19 November 2026