```

Uber is notifying close to 150,000 people after attackers used login credentials belonging to one of its technology suppliers to reach personal data held in an external environment — a breach that turns the familiar third-party risk problem into something sharper for teams now buying into AI SaaS.

According to Security Affairs, which published the disclosure on 9 October 2026, the compromised credentials belonged to Fireworks AI, an AI infrastructure provider. The attackers used them to access an Uber external environment and take personal information, including driver's licence details, belonging to nearly 150,000 Uber users. Uber itself disclosed that "on October 9, 2026, a service provider" was the entry point for the intrusion.

Two caveats belong up front. The identification of Fireworks AI as that service provider rests on Security Affairs' reporting rather than an independently verified attribution, and it should be treated as such. How the credentials were captured — phishing, credential stuffing, a compromise of the vendor's own systems, or something else entirely — is not yet clear, and the source does not report any publication or sale of the data.

What makes this incident distinctive is not the scale but the vector. Uber did not suffer a failure of its own authentication controls; a credential issued to a third party was enough to open a path to a store of user records. For organisations deploying AI services — where providers routinely require access to endpoints, data feeds or integration credentials as part of onboarding — the incident reads as a straightforward reminder that vendor access is standing access, and it should be governed like any other.

It also raises a data-minimisation question worth asking before, not after: why did an external environment reachable by provider-held credentials contain driver's licence numbers at all? Driver's licence data is a high-value target precisely because it is difficult to rotate, unlike a password. Any environment exposed to third-party credentials is a candidate for holding the least sensitive data the business function will tolerate.

The breach arrives amid wider scrutiny of how AI vendors are assessed. Procurement teams that would not dream of granting an unvetted contractor standing access to production systems often wave through AI providers with a security questionnaire and a signature. The asymmetry is hard to justify.

Takeaways for security teams

  • Inventory AI vendors with payment-processor rigor. Treat every AI SaaS provider as a party with reach into your environment, and keep a current register of who holds which credentials.
  • Treat provider access as production access. Scope it, time-box it, and monitor it with the same tooling used for privileged internal accounts.
  • Write notification timelines into contracts. Uber's disclosure turned on a service provider's involvement; the speed of your own breach response depends on how fast the vendor tells you.
  • Assume credential reuse. A provider credential compromised elsewhere may already be valid against your integrations — rotate, and check for reuse across environments.

For Hong Kong organisations, the incident lands squarely on the due-diligence expectations that sit alongside the Personal Data (Privacy) Ordinance's data user obligations: when you engage a processor, the responsibility for proper handling of personal data does not outsource with it. Contractual assurances, access restrictions and evidence of the vendor's own controls are the practical means of discharging that responsibility — and an AI provider with programme-level access to your environment deserves the same scrutiny as any other processor touching customer records.

Whether this becomes a one-off or a template depends largely on how quickly organisations stop treating AI procurement as a productivity decision rather than a security one.


```

Uber 正在通知接近 15 萬名用戶,此前攻擊者利用其一家技術供應商的登入憑證,存取了儲存在外部環境中的個人資料 —— 這宗外洩事件將人們熟悉的第三方風險問題,對正大舉採購 AI SaaS 的團隊而言變得更為尖銳。

據 Security Affairs 報道(該機構於 2026 年 10 月 9 日披露事件),遭洩露的憑證屬於 AI 基礎設施供應商 Fireworks AI。攻擊者利用這些憑證存取 Uber 的一個外部環境,竊取近 15 萬名 Uber 用戶的個人資料,其中包括駕駛執照資料。Uber 自身披露稱,「2026 年 10 月 9 日,一家服務供應商」是入侵的起點。

有兩點須先說明。將 Fireworks AI 認定為該服務供應商,是基於 Security Affairs 的報導,而非經獨立核實的歸因,應按此理解。憑證如何被截取 —— 是 phishing、credential stuffing、供應商自身系統被入侵,抑或其他方式 —— 目前尚不清楚,而相關消息來源亦未報道有任何資料被公開或出售。

令這宗事件獨特的不是規模,而是攻擊向量(vector)。Uber 自身的authentication controls 並未失效;一張發予第三方的憑證,已足以打通通往用戶記錄庫的途徑。對於部署 AI 服務的機構 —— 此類供應商通常會在入駐(onboarding)過程中要求存取端點、data feeds 或 integration 憑證 —— 這宗事件可謂一個直接的提醒:供應商的存取權即是長期存取權,應如同其他存取權一樣受到管治。

事件同時引出一個應在事前而非事後追問的 data minimisation 問題:為何一個可由供應商所持憑證存取的外部環境,竟然會存放駕駛執照號碼?駕駛執照資料之所以是高價值目標,正因為它與密碼不同,難以輪換(rotate)。任何暴露於第三方憑證的環境,都應只存放該業務環節所能容忍的最不敏感資料。

這宗外洩事件發生之際,正值外界對 AI 供應商評估方式的檢視升溫。採購團隊從不敢讓未經審查的外判商長期存取生產系統,卻往往憑一份 security questionnaire 和一個簽名便放行 AI 供應商。這種不對稱難以自圓其說。

安全團隊要點

  • 以處理付款處理商的嚴謹程度盤點 AI 供應商。 將每一家 AI SaaS 供應商視為有能力接觸你環境的一方,並維護一份最新名冊,記錄誰持有哪些憑證。
  • 將供應商存取視為生產環境存取。 為其界定範圍(scope)、設定時限(time-box),並以監察內部特權帳戶的同一套工具進行監控。
  • 在合約中訂明通知時限。 Uber 的披露取決於服務供應商是否牽涉其中;你自身外洩應變的速度,取決於供應商通知你的快慢。
  • 假設憑證會被重用。 供應商憑證若在其他地方被入侵,可能對你的 integration 仍然有效 —— 應立即輪換(rotate),並檢查各環境之間是否存在重用情況。

對香港機構而言,這宗事件正好落在與《個人資料(私隱)條例》資料使用者責任並列的盡職審查期望之上:當你聘用一個處理者(processor),妥善處理個人資料的責任不會隨之外判。合約上的保證、存取限制,以及供應商自身控制措施的證明,乃是履行該責任的實際手段 —— 而一家擁有程式層級存取權接觸你環境的 AI 供應商,理應獲得與任何接觸客戶記錄的處理者同等的審查。

這宗事件會否成為孤例或範本,很大程度取決於機構何時停止將 AI 採購視為生產力決定,而非安全決定。

新聞來源 / Original News Source