Having an officially sanctioned AI workspace does not, by itself, stop staff from typing company information into personal chatbot accounts. That "approved ≠ controlled" gap is the focus of guidance from NordLayer, reported by BleepingComputer on 16 October, which lays out how IT administrators can detect unauthorised generative-AI use and keep sensitive data out of consumer-grade tools.
The scenario is familiar to many IT departments: an organisation licenses an enterprise AI workspace with appropriate data-handling terms, assumes the problem is solved, and discovers that employees still default to the personal ChatGPT accounts they already have on their phones and home browsers. The sanctioned workspace creates what amounts to a false sense of assurance. Data governance depends on what staff actually use day to day, not on what procurement has signed off.
The exposure is concrete. When an employee pastes a customer list, source code, internal financials, or contract language into a consumer account, the information leaves the corporate environment entirely — outside the organisation's audit trail, retention controls, and access management. NordLayer recommends treating this as an investigation and remediation exercise rather than a policy memo.
A further wrinkle concerns model training. Under OpenAI's data-controls documentation, content entered into consumer ChatGPT can be used to help improve its models unless the user actively disables that option in their account settings; content sent through the business and API tiers is not used for training by default. The opt-out, however, is a per-account setting — one that individual employees control and IT teams have no reliable way to verify across the workforce. That asymmetry is one reason consumer accounts are a poor fit for corporate data, regardless of a user's intentions.
A practical playbook for administrators
The guidance points IT teams toward a layered response, sequenced so that investigation precedes enforcement:
- Analyse network traffic for connections to consumer AI domains from managed devices, including traffic on mobile devices enrolled in the organisation's mobile-device management (MDM) programme.
- Inventory browser extensions and applications that interact with AI services, looking for unofficial plugins, wrappers, and integrations that may route company data externally.
- Review single sign-on and identity logs for signs that personal email addresses are being used in conjunction with work resources.
- Tune existing data-loss-prevention rules with prompts and patterns tailored to likely sensitive content — pasted code blocks, spreadsheets, and internal terminology.
- Block or restrict high-risk services at the network or policy level once the investigation has established scope, rather than acting on assumptions.
The sequencing matters: blanket blocks imposed before an inventory tend to drive usage further underground and teach staff to route around controls on unmanaged networks and devices — reducing visibility without reducing risk.
Regulatory expectations in Hong Kong
Locally, the Privacy Commissioner for Personal Data (PCPD) published guidance on generative AI in 2024, setting out expectations for organisations deploying these technologies — including principles around data minimisation, transparency, and human oversight. The guidance does not specifically prohibit consumer chatbot use, but it does place responsibility on organisations to understand and manage how personal data is handled across the tools their staff use. IT teams that cannot answer the question "where did that data go?" may find it difficult to demonstrate compliance with those principles.
Beyond detection: governance, not prohibition
NordLayer's closing recommendation is worth noting for what it avoids. Rather than advocating blanket bans on generative AI — an approach widely regarded as unworkable given how embedded these tools have become — it argues for pairing a sanctioned tool with a clear acceptable-use policy. The two together give staff a legitimate, governed alternative and give IT a basis for enforcement.
For Hong Kong IT professionals, the underlying message is that shadow AI is a governance problem before it is a technical one. Tools can detect and restrict, but only policy, training, and an approved option that people actually want to use will close the gap between what the organisation believes is happening and what is actually happening.
The guidance summarised here was published by NordLayer and reported by BleepingComputer on 16 October.
擁有一個官方批准的 AI 工作平台，本身並不能阻止員工將公司資料輸入個人 chatbot 帳戶。這個「已批准 ≠ 受控制」的落差，正是 NordLayer 最新指引的重點所在。BleepingComputer 於 10 月 16 日報道了該指引，內容詳述 IT 管理員如何偵測未經授權的生成式 AI 使用情況，以及如何防止敏感資料流入消費級工具。
許多 IT 部門對以下場景並不陌生：機構購買了具備適當資料處理條款的企業 AI 工作平台，以為問題已經解決，卻發現員工仍然慣用他們手機及家用瀏覽器上已有的個人 ChatGPT 帳戶。獲批准的平台反而造成一種虛假的安全感。數據管治取決於員工日常實際使用的工具，而非採購部門簽核批准了甚麼。
風險是具體實在的。當員工將客戶名單、source code、內部財務資料或合約條文貼入消費級帳戶時，該資料便完全脫離企業環境 —— 不在機構的 audit trail、資料保留控制及存取管理範圍之內。NordLayer 建議將此視為一項調查及補救工作，而非僅僅發出一份政策文件。
另一個複雜情況涉及模型訓練。根據 OpenAI 的資料控制文件，輸入消費級 ChatGPT 的內容可用於改進其模型，除非用戶主動在其帳戶設定中關閉該選項；而透過商業版及 API 層級發送的內容，預設不會用於訓練。然而，「選擇退出」只是一個按帳戶設定的選項 —— 由個別員工自行控制，IT 團隊無法可靠地在全體員工範圍內核實是否已啟用。這種不對稱，正是消費級帳戶不適合處理企業資料的原因之一，與用戶的意圖無關。
給管理員的實務操作手冊
該指引建議 IT 團隊採取分層回應，並按次序進行，先調查、後執行：
- 分析網絡流量，留意受管理設備連接消費級 AI 網域的記錄，包括已納入機構流動設備管理（MDM）計劃的流動設備流量。
- 盤點與 AI 服務互動的瀏覽器外掛程式及應用程式，找出可能將公司資料對外轉送的非官方 plugin、wrapper 及整合工具。
- 檢視 single sign-on 及身份識別記錄，查看是否有使用個人電郵地址配合工作資源的跡象。
- 調整現有的數據防外洩（DLP）規則，加入針對可能敏感內容的 prompt 及模式 —— 例如貼上的 code block、試算表及內部專有用詞。
- 在網絡或政策層面封鎖或限制高風險服務，前提是調查已確定影響範圍，而非基於假設行事。
次序非常重要：在完成盤點之前實施全面封鎖，往往只會令使用情況轉入更隱蔽的途徑，員工學會透過不受管理的網絡及設備繞過控制 —— 結果是可視性下降，風險卻沒有降低。
香港的監管期望
在香港，個人資料私隱專員公署（PCPD）於 2024 年發布了生成式 AI 指引，訂明機構採用此類技術時的期望 —— 包括數據最小化、透明度及人為監督等原則。該指引並非明文禁止使用消費級 chatbot，但確實要求機構了解並管理員工所用各項工具如何處理個人資料。無法回答「那些資料去了哪裡？」的 IT 團隊，可能難以證明自己符合相關原則。
超越偵測：管治而非禁止
NordLayer 最後的建議值得留意之處，在於它避開了甚麼。它並非主張全面禁止生成式 AI —— 鑑於這類工具已根深蒂固，該做法普遍被認為不可行 —— 而是主張將獲批准的工具與清晰的可接受使用政策配合實施。兩者結合，為員工提供一個正當、受管治的替代方案，亦為 IT 提供執行的依據。
對香港 IT 專業人士而言，核心信息是：shadow AI 首先是管治問題，然後才是技術問題。工具可以偵測和限制，但只有政策、培訓，以及一個員工真正願意使用的獲批准選項，才能彌合機構所相信的情況與實際發生的情況之間的落差。
上述指引由 NordLayer 發布，並由 BleepingComputer 於 10 月 16 日報道。