Matt Robb intended to sell his computer keyboard, so he posted the item information on Facebook Marketplace and, at the end of last month, conducted an experiment in which he let Muse, Meta's newly launched AI agent, take charge of the entire process.
馬特·羅布打算出售他的電腦鍵盤,於是在Facebook Marketplace上刊登了商品資訊,並在上個月末進行了一項實驗,讓Meta新推出的人工智慧代理Muse主導整個過程。
Muse began to operate.
Muse開始運作。
However, within less than a day, Robb received an alarming update: a buyer he had never met showed up outside his front door late at night.
然而不到一天,羅布便收到了一則令人警惕的更新:一名他素不相識的買家在深夜時分出現在他家門外。
What exactly had happened?
究竟是怎麼回事?
Muse had closed the deal on its own and, after agreeing to an in-person handover, shared Robb's home address.
Muse自行達成了交易,並在同意面交之後分享了羅布的家庭住址。
When the buyer arrived, it even automatically sent a message from his account saying "Yes, I'm here!" — and Robb was completely unaware of all of this.
當買家抵達時,它甚至從他的帳號自動發送了一則「沒錯,我在這裡!」的訊息——而這一切羅布全然不知情。
"Not only did it give the other person my address when they asked and accept the offer, but it was almost as if it were using me to tease the guy (while I wasn't there)," Rob said.
「它不僅在對方詢問時給了對方我的地址、接受了出價,而且幾乎就像是在(我不在時)拿我來逗那個傢伙,」羅布說道。
He is a content creator based in Toronto, and he also described this experience in detail in his social media post, which has now been widely shared.
他是多倫多的一名內容創作者,也在他那則如今被廣泛分享的社群媒體貼文中詳細描述了這段經歷。
Many people have already integrated AI agents into their online lives and have praised personalized assistants like Muse, saying they can easily handle everyday tasks — from booking dinner reservations and planning trips to organizing email inboxes.
許多人已將AI代理融入自己的線上生活,並對Muse這類個人化助理讚譽有加,稱其能輕鬆處理日常事務——從預訂晚餐位子、規劃旅行,到整理電子郵件收件匣。
However, like other AI bots, these agents' performance depends on the information they can access.
然而,與其他AI機器人一樣,這些代理的表現取決於它們能取得的資訊。
The more they know about the user, the better they perform — perhaps including knowledge about that user's friends, family, or even colleagues.
它們對使用者了解得越多,表現就越好——或許也包括對該使用者的朋友、家人甚至同事的了解。
This raises privacy concerns.
這引發了隱私疑慮。
Rob and the buyer eventually reached an understanding, and he also learned from Meta that the sharing of his address ultimately stemmed from the permission he had originally granted Muse to access his Marketplace profile.
羅布與買家最終達成諒解,他亦從Meta得知,分享其地址歸根究柢源於他當初首次授予Muse存取其Marketplace個人檔案的權限。
However, this incident revealed the real-world risks that sharing large amounts of personal information with such technology can give rise to.
然而,這起事件揭示了與該技術分享大量個人資訊可能引發的真實世界風險。
"Using an AI agent is like handing someone your keys and a to-do list," said Veronica Rivera, an assistant professor at Georgia Tech's School of Cybersecurity and Privacy.
「使用AI代理就像把你的鑰匙和一份待辦清單交給別人,」喬治亞理工學院網路安全與隱私學院助理教授維羅妮卡·里維拉說。
"How much they can help you, and how much risk that creates, depends on which keys you hand over."
「它們能幫上多少忙、因此帶來多大風險,取決於你交出了哪些鑰匙。」
Since its launch in September, downloads of Meta's Muse app have risen sharply, and competition is only expected to intensify.
自九月推出以來,Meta的Muse應用程式下載量已大幅攀升,而競爭預料只會愈演愈烈。
OpenAI has also recently launched its own series of "always-on" AI assistants called dots, which can be integrated into ChatGPT as well as workplace platforms such as Slack and Microsoft Teams.
OpenAI近日亦推出自家一系列「始終開啟」的AI助手,名為dots,可整合至ChatGPT,以及Slack和Microsoft Teams等工作場所平台。
Early generative AI chatbots, such as the standard version of ChatGPT, responded to specific prompts; at the user's request, they could write articles, generate entirely new images or music, and even write code.
早期的生成式AI聊天機器人,諸如標準版的ChatGPT,乃是針對特定提示作出回應;在使用者的要求下,它們得以撰寫文章、生成嶄新的圖像或音樂,甚至編寫程式碼。
Meta's Muse and OpenAI's dots go a step further, capable of carrying out everyday tasks on a person's behalf.
Meta的Muse與OpenAI的dots更進一步,能代替個人執行日常任務。
AI agents are powered by the same large language models as chatbots, but they operate in a loop, which enables an agent to take action on its own initiative even when the user is not watching.
AI代理由與聊天機器人相同的大型語言模型驅動,惟其以迴圈方式運行,從而使代理得以主動採取行動,即使使用者並未在旁留意。
"What makes these AI agents truly appealing is that they are more useful than ever before," Rivera said.
「讓這些AI代理真正吸引人的一點是,它們前所未有地有用,」里維拉說道。
She is also a trustworthy AI researcher at the Computing Research Association, which is funded by Microsoft.
她同時也是微軟資助的計算研究協會可信AI研究員。
"If you give an agent access to your credit card, that agent can now instantly obtain all your historical transaction records... they will absorb all of this data in less than a few seconds," Bagri said.
「如果你把信用卡的存取權限提供給一個代理,該代理現在就能立即取得所有歷史交易紀錄……它們會在不到幾秒內吸收所有這些資料,」巴格里說。
Although users can later revoke access, Bagri says that data an agent has already seen "cannot be taken back," and how developers store historical data may also be unclear.
雖然使用者之後可以撤回存取權,但巴格里表示,代理已經看過的資料「無法收回」,而開發者如何儲存歷史資料也可能模糊不清。
Both Meta and OpenAI have outlined their relevant privacy policies for the new AI agents running on their respective dedicated cloud computers.
Meta和OpenAI皆已就其各自專屬雲端電腦上運行的新型AI代理,概述了相關隱私政策。
While users can turn off the memory-sharing function between dots and ChatGPT, the information the agents previously received will not be automatically deleted as a result.
使用者固然可以關閉dots與ChatGPT之間的記憶共享功能,然而代理先前所接收的資訊並不會因此自動刪除。
OpenAI offers users the option to request access to, deletion of, or correction of their personal information, although the relevant rights may vary depending on where they live.
OpenAI提供使用者請求存取、刪除或更正個人資訊的選項,惟相關權利可能因居住地點而有所差異。
Meta's Muse also gives users the option to delete message history and other information.
Meta的Muse亦為使用者提供刪除訊息歷史及其他資訊的選項。
In everyday use, the agent asks users for access permissions — Meta says that Muse learns about users and the people they interact with "based on public information and the content you choose to share."
在日常使用中,該代理會向使用者請求存取權限——Meta表示,Muse「根據公開資訊以及你選擇分享的內容」來了解使用者及其互動對象。
However, the company says that this information remains safe and secure for every user.
然而,該公司表示,這些資訊對每位使用者均維持安全無虞。
In Rob's case, he learned that the leak of his address ultimately came down to the "always allow" permission he had granted when he first gave Muse access to his Facebook Marketplace profile; he had shared his in-person pickup address in that profile, and this was merely part of a broader template intended for the agent to respond to potential buyers.
在羅布的案例中,他了解到地址之所以外洩,歸根究柢在於他當初首次讓Muse存取其Facebook Marketplace個人檔案時所授予的「永遠允許」權限;他在該檔案中分享了自己的面交地址,而這僅是供代理回應潛在買家的更廣泛範本的一部分。
However, he warned that others might likewise misunderstand the scope of that access permission — and said he had initially assumed that, regardless, the AI agent would check with him before sharing that information or accepting an offer.
但他警告,其他人也可能同樣誤解該存取權限的範圍——並表示他起初以為,無論如何AI代理都會在分享該資訊或接受出價之前先與他確認。
His AI agent was not entirely error-free in carrying out its remaining instructions either.
他的AI代理在執行其餘指令時亦非毫無差錯。
For another item Rob was selling, Muse appeared to suggest that he would accept a price $100 lower than the minimum he had previously set.
對於羅布正在出售的另一件商品,Muse似乎暗示他會接受比先前設定最低價低100美元的價格。
When contacted by reporters, Meta acknowledged the agent's pricing error but declined to comment further.
記者聯繫時,Meta承認了該代理的定價錯誤,但未進一步置評。
Rob later largely disconnected Muse from his application.
羅布後來大致上已將Muse與他的應用程式斷開連接。
In an interview earlier this month, he said the situation could have been far worse, especially "if a stranger had shown up at the door of someone more vulnerable".
在本月早些時候的一次採訪中,他表示情況原本可能糟得多,尤其是「如果某個陌生人出現在更為脆弱的人家門口」。
Nevertheless, Rob said he expects these agents to be fine-tuned, and that he would consider reactivating Muse in the future.
儘管如此,羅布表示他預期這些代理將經過微調,而他未來也會考慮重新啟用Muse。
Since then, he has been testing another AI agent — Grok Bot from Elon Musk's X.
此後,他一直在測試另一個AI代理——伊隆·馬斯克旗下X的Grok Bot。
Other tech companies have already become wary of the massive influx of AI agents.
其他科技公司對於AI代理的大量湧入已有所警惕。
Earlier this month, Apple said it would introduce new controls for its Mac operating system to address AI agents gaining broader access to apps such as email, messages and even browsing history when users are not "fully informed and aware."
本月早些時候,蘋果表示將為其Mac作業系統引入新的管控措施,以應對AI代理在使用者並非「完全知情與理解」的情況下,更廣泛地存取郵件、訊息乃至瀏覽紀錄等應用程式。
Apple did not name any developer or agent, but the update was released after Jason Aten, a columnist for Inc. magazine, reported that Muse viewed notifications on his Mac (including Messages) without him having granted the agent specific access rights.
蘋果並未指明任何開發者或代理的名稱,但這項更新是在《Inc.》雜誌專欄作家傑森·阿滕報導Muse在他未授予該代理特定存取權的情況下查看他Mac上的通知(包括Messages)之後發布的。
Meta spokesperson Andy Stone insisted that such integrations are "entirely opt-in," and wrote on social media last month that users must enable both Apple's security feature "Full Disk Access" and the Messages connector through Muse; Aten, for his part, said he had not granted the agent these permissions.
Meta發言人安迪·史東堅稱,這類整合「完全是選擇加入的」,並於上個月在社群媒體上寫道,使用者必須同時啟用蘋果的安全功能「完整磁碟存取」以及透過Muse的Messages連接器;阿滕則表示,他並未授予該代理這些權限。
Even those who have not jumped on the AI agent bandwagon are not necessarily unaffected or protected.
即便是那些並未加入AI代理熱潮的人,也未必就能置身事外、受到保障。
Consumers who have not downloaded AI agents may still end up interacting with those who have.
未下載AI代理的消費者,仍可能與已下載者產生互動。
And beyond the confusion that may arise when AI sends messages on users' behalf, experts also warn of a knock-on effect on privacy.
而除了AI代替使用者發送訊息時可能引發的混淆之外,專家亦警告隱私方面將產生連鎖效應。