Toronto content creator Matt Robb was selling a keyboard on Facebook Marketplace. He let Muse, Meta's new AI agent, handle the deal.
多倫多內容創作者馬特·羅布在Facebook Marketplace上賣鍵盤,並讓Meta新推出的AI代理Muse主導交易。
In less than a day, a buyer he did not know showed up outside his home late at night.
不到一天,一名他不認識的買家就在深夜出現在他家門外。
It turned out that Muse made the deal on its own. It agreed to meet in person and shared his home address.
原來Muse自行達成交易,同意面交並分享了他的住址。
It also sent a message by itself saying "Yes, I'm here!" Robb had no idea any of this was happening.
它還自動發訊息說「沒錯,我在這裡!」,羅布全程不知情。
Many people praise personalized assistants like Muse for handling everyday tasks, such as booking restaurants, planning trips and organizing inboxes.
許多人讚賞Muse這類個人化助理能處理日常事務,例如訂餐廳、規劃旅行和整理收件匣。
But experts point out that these agents perform better the more they know about users. This also raises privacy concerns.
但專家指出,這些代理對使用者了解越多,表現就越好,這也引發隱私疑慮。
Georgia Tech assistant professor Rivera said using an AI agent is like handing someone your keys and to-do list.
喬治亞理工學院助理教授里維拉說,使用AI代理就像把鑰匙和待辦清單交給別人。
The risk depends on which keys you hand over.
風險取決於你交出了哪些鑰匙。
Meta's Muse has seen downloads soar since it launched in September.
Meta的Muse自九月推出以來下載量飆升。
OpenAI has also launched an "always-on" assistant called dots. It can be integrated into ChatGPT and work platforms.
OpenAI也推出可整合進ChatGPT和工作平台的「始終開啟」助手dots。
University of Toronto professor Bagri said that once an agent gets user information, the data it has seen "cannot be taken back." This is true even if permission is withdrawn later.
多倫多大學教授巴格里表示,代理一旦取得使用者資訊,即使之後撤回權限,已看過的資料也「無法收回」。
Robb later learned that the address leak came from the "always allow" permission he had granted.
羅布後來得知,地址外洩源於他當初授予的「永遠允許」權限。
He had assumed the agent would check with him before sharing information.
他原以為代理分享資訊前會先與他確認。
Meta admitted that Muse wrongly suggested he would accept a $100 discount on another item.
Meta承認Muse在另一件商品上錯誤暗示他願降價100美元。
Robb has mostly disconnected Muse, but he said he might turn it on again in the future.
羅布已大致斷開Muse,但表示未來可能重新啟用。
Apple also announced new controls for the Mac this month to address the access risks of AI agents.
蘋果本月也宣布為Mac加入新控制措施,應對AI代理的存取風險。
Rivera warned that when an agent reads your messages, it also reads what friends and family sent you. That is true even if they never agreed.
里維拉警告,代理讀取你的訊息時,也讀取了親友寄給你的內容,即使他們從未同意。