Matt Robb is a content creator in Toronto. He sells keyboards on Facebook Marketplace.
多倫多內容創作者馬特·羅布在Facebook Marketplace上賣鍵盤。
He let Muse, Meta's new AI agent, lead the deal.
他讓Meta新推出的AI代理Muse主導交易。
In less than a day, a stranger came to his door. It was 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 by itself.
原來Muse自行達成交易。
It agreed to meet in person and shared his home address.
它同意面交,還分享了他的住址。
It also sent a message by itself: "Yes, I am here!"
它還自動發訊息說「沒錯,我在這裡!」。
Robb did not know about any of this.
羅布全程不知情。
Many people like Muse and other personal helpers. They can do everyday tasks.
許多人讚賞Muse這類個人化助理能處理日常事務。
For example, they book restaurants, plan trips, and clean your inbox.
例如訂餐廳、規劃旅行和整理收件匣。
But experts say these agents work better when they know the user more.
但專家指出,這些代理對使用者了解越多,表現就越好。
This also raises privacy worries.
這也引發隱私疑慮。
Rivera is a professor at Georgia Tech. He says using an AI agent is like giving someone your keys. And your 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 the keys you give away.
風險取決於你交出了哪些鑰匙。
Meta's Muse has had a big rise in downloads since it launched in September.
Meta的Muse自九月推出以來下載量飆升。
OpenAI also launched an "always on" helper called dots. It can work inside 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 this: once an agent gets user information, it cannot be taken back. Even if you remove permission later, the data it saw is "gone for good."
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」。
Rob later learned the address leak came from the "always allow" permission he gave.
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 thought 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 Muse wrongly said he would lower the price by $100 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。
Rob has mostly disconnected Muse, but he says he may turn it on again.
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 Mac this month. They are for AI agent access risks.
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 warns that an agent reads your messages. It also reads what friends and family sent you.
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。
Even if they never agreed to it.
即使他們從未同意。