Opening a shared ChatGPT link could have let an attacker run their own code inside your logged-in browser session — and read the private conversations stored there — according to researchers at the security firm Parameter, as reported by The Hacker News.
The exposure was direct and personal: anyone who clicked a booby-trapped share link while signed in to ChatGPT handed the attacker the ability to execute code in a context that already carried their session, and with it access to their chat history. For a tool that increasingly holds drafts, source code, business proposals and internal analysis, a link shared casually in a chat window or email thread was enough to put that material within reach of someone else.
The fault, according to the report, sat in the sharing and session layer rather than in the language model's own processing. In other words, this was a familiar class of web trust-boundary bug wearing a new address: an untrusted link meeting an authenticated user, with nothing sufficient in between. The novelty is not the weakness itself but its placement — directly in front of the conversational interfaces companies are folding into everyday work.
Parameter disclosed the flaw through OpenAI's bug bounty programme on 18 October. OpenAI patched it eight days later and classified it as P1, the most serious tier in its own rating scale — a signal that the company viewed the issue as more than routine housekeeping.
The timing matters. ChatGPT share links are routinely circulated in group chats, tickets and documentation, precisely because sharing is the feature's purpose. That makes the attack surface social rather than technical: a convincing link costs an adversary nothing, and the recipient's only defence is the habit of checking where it actually points.
What the incident does not tell us is just as important as what it does. The source material does not indicate whether the flaw was exploited before the patch landed, nor does it describe how the fix was delivered to users or whether a working proof-of-concept was released. Those details remain unstated, and readers should treat them as unknown rather than reassuring.
For Hong Kong workplaces, where link-sharing across messaging apps and internal channels is routine, the practical response is proportionate rather than dramatic. Treat share links from AI services with the same caution as any other unsolicited URL: confirm the domain before clicking, avoid opening unfamiliar links while signed in to accounts holding sensitive material, and keep credentials, keys and confidential data out of chat sessions entirely. None of this requires new tooling — only the ordinary link hygiene that security teams already ask of email.
OpenAI's eight-day turnaround and P1 rating show the vendor treated the report seriously. The residual risk now sits with habits, not patches: as AI assistants become the default place to work through problems, the boundaries around them deserve the same scrutiny as any other production system.
據安全公司 Parameter 研究人員透露，並由 The Hacker News 報道，打開一個共享的 ChatGPT 鏈接，可能令攻擊者在你已登入的瀏覽器會話中執行任意代碼，並讀取儲存在該處的私人對話內容。
這次漏洞的影響直接而切身：任何人在登入 ChatGPT 狀態下點擊了被動過手腳的分享鏈接，等同把執行代碼的能力交到攻擊者手上，而且是在已載有其會話憑證的上下文環境中執行，從而取得其聊天紀錄的存取權。對於一個越來越多用於保存草稿、源代碼、商業方案及內部分析的工具而言，隨手在聊天視窗或電郵中分享的一個鏈接，已足以令這些材料落入他人可及的範圍。
報告指出，漏洞出自分享及會話層，而非語言模型本身的處理環節。換言之，這是一類常見的網絡信任邊界（trust-boundary）漏洞，只是換上了新的地址：一個不受信任的鏈接遇上一個已通過認證的用戶，中間缺乏足夠的防護。其新意並不在於漏洞本身，而在於它的位置——直接處於企業正逐步整合到日常工作的對話式界面之前。
Parameter 於 10 月 18 日透過 OpenAI 的 bug bounty 計劃披露此漏洞。OpenAI 在八日後完成修復，並將其列為 P1，即自身評級中最嚴重的級別——顯示該公司視此問題非同一般的例行修補。
時機因素至關重要。ChatGPT 分享鏈接本來就經常在群組對話、工單及文件中流傳，因為分享正是該功能的設計目的。這使得攻擊面屬社會層面而非技術層面：一條看似可信的鏈接對攻擊者而言毫無成本，收件人唯一的防線，就是養成核實鏈接實際指向的習慣。
這起事件「沒有告訴我們的事」，與它所揭示的同樣重要。原始資料並未說明該漏洞在補丁推出前是否曾被利用，亦未描述修復方式如何交付予用戶，以及是否已有可運行的 proof-of-concept 被公開。這些細節均未有交代，讀者應視之為未知，而非可以安心。
對於香港職場而言，在各類即時通訊應用程式及內部渠道之間分享鏈接屬常態操作，務實的應對方式應當恰如其分，毋須過度渲染。處理 AI 服務發出的分享鏈接，應與其他未經主動索取的 URL 一樣保持警惕：點擊前先核實域名；避免在登入載有敏感材料的帳戶時打開陌生鏈接；以及切勿在聊天會話中處理任何憑證、密鑰及機密資料。這些做法無須引入新工具——只需執行安全團隊早已對電郵提出的日常鏈接衞生習慣。
OpenAI 八日內完成修復並將漏洞列為 P1，顯示供應商對報告予以重視。現時的殘餘風險在於習慣，而非補丁：隨著 AI 助手成為處理問題的預設工作方式，圍繞它們的邊界理應獲得與其他生產系統同等的審視。