Scarcely had Muse been set in motion when, within less than a day, Robb received an alarming update: a buyer he had never met turned up outside his front door in the dead of night.
Muse甫一啟動運作,羅布旋即於不到一日之內接獲一則令人警覺的更新:一名他素不相識的買家竟於深夜時分現身其家門之外。
What lay behind this turn of events?
箇中緣由何在?
Muse had clinched the deal of its own accord and, having agreed to an in-person handover, disclosed Robb's address.
Muse自行促成了交易,並於同意面交之後披露了羅布的住址。
When the buyer arrived, it even went so far as to fire off a message from his account of its own volition — "Yep, I'm here!" — and Robb was kept entirely in the dark throughout.
當買家抵達之際,它甚至逕自從其帳號自動發送了一則「沒錯,我在這裡!」的訊息——而這一切,羅布全然蒙在鼓裡。
"Not only did it give the guy my address when he asked for it and accept the offer, but it was almost as if it was teasing the guy with me 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 recounted 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 welcomed AI agents into their online lives, lavishing praise on personalised assistants such as Muse, which they say handle everyday chores with ease — from securing dinner reservations and planning trips to tidying up their email inboxes.
許多人已將AI代理迎入自己的線上生活,並對Muse這類個人化助理讚不絕口,稱其能輕鬆處理日常事務——從搶訂晚餐位子、規劃旅行,到整理電子郵件收件匣。
Yet, just like any other AI bot, these agents are only as good as the information they can access.
然而,與其他AI機器人如出一轍,這些代理的優劣全然取決於它們所能取得的資訊。
The more thoroughly they understand the user, the better they perform — perhaps including an understanding of that user's friends, family and even colleagues.
它們對使用者了解得越透徹,表現便越臻完善——或許也包括對該使用者的朋友、家人乃至同事的了解。
This fact gives rise to privacy concerns.
此一事實引發了隱私疑慮。
Rob and the buyer eventually reached an understanding; he also learned from Meta that the sharing of his address ultimately stemmed from the permissions he had granted when he first gave Muse access to his Marketplace profile.
羅布與買家最終達成了諒解;他亦從Meta獲悉,分享其地址一事,歸根究柢源於他當初首次授予Muse存取其Marketplace個人檔案時所賦予的權限。
Nevertheless, this episode lays bare the real-world risks that can arise from sharing substantial amounts of personal information with such technology.
然而,這起事件揭示了與該技術分享大量個人資訊所可能引發的真實世界風險。
"Using an AI agent is tantamount to entrusting someone with your keys along with a to-do list," said Veronica Rivera, an assistant professor in the School of Cybersecurity and Privacy at Georgia Tech.
「使用AI代理,無異於將你的鑰匙連同一份待辦清單一併託付於人,」喬治亞理工學院網路安全與隱私學院助理教授維羅妮卡·里維拉表示。
"How great the benefits are—and, by extension, how high the attendant risks—depends entirely on which keys you actually hand over."
「其助益之大小,乃至隨之而來的風險之高低,皆取決於你究竟交出了哪些鑰匙。」
Since its launch in September, downloads of Meta's Muse app have soared; and the competition is only set to intensify.
自九月面世以來,Meta旗下Muse應用程式的下載量便急遽攀升;而競爭態勢只會愈演愈烈。
OpenAI has likewise recently rolled out its own suite of "always-on" AI assistants, dubbed 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、Microsoft Teams等工作場所平台之中。
Earlier generative AI chatbots—such as the standard version of ChatGPT—operated on nothing more than a mechanism of responding to specific prompts; once they received a user's request, they could write essays, generate new images or music, and even write code.
較早的生成式AI聊天機器人——諸如標準版的ChatGPT——其運作機制無非是針對特定提示作出回應;一旦接獲使用者的要求,它們便得以撰寫文章、生成新圖像或音樂,甚而編寫程式碼。
Meta's Muse and OpenAI's dots go a step further, possessing the capacity to perform everyday tasks on an individual's behalf.
Meta的Muse與OpenAI的dots則更進一步,具備代替個人執行日常任務的能力。
Although AI agents are powered by the same large language models as chatbots, they operate in a loop, thereby enabling the agent to take the initiative and act even when the user is not watching over it.
AI代理雖由與聊天機器人相同的大型語言模型所驅動,卻以迴圈方式運行,從而使代理得以主動採取行動,即便使用者並未在旁留意。
"If you grant an agent access to your credit card, that agent can instantly obtain all historical transaction records... they will absorb all of this data within mere seconds," Bagri said.
「倘若你將信用卡的存取權限授予一個代理,該代理便得以即時取得所有歷史交易紀錄……它們會在寥寥數秒之內將這些資料盡數吸收,」巴格里說道。
Although users may later revoke access, Bagri says that data the agent has already perused is "impossible to reclaim," and precisely how developers store historical data is likely to remain murky.
雖則使用者日後可撤回存取權,然巴格里表示,代理業已閱覽之資料「無從收回」,而開發者究竟如何儲存歷史資料,亦恐晦暗不明。
Both Meta and OpenAI have outlined their respective privacy policies for their new AI agents that run on dedicated cloud computers.
Meta與OpenAI皆已就其各自於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 indeed disable the memory-sharing feature between dots and ChatGPT, the information the agent previously received will not be automatically deleted as a result.
使用者固然得以關閉dots與ChatGPT之間的記憶共享功能,惟代理先前所接收的資訊並不會因此自動刪除。
Although OpenAI offers users the option to request access to, deletion of, or correction of their personal information, such rights may vary depending on where one resides.
OpenAI雖提供使用者請求存取、刪除或更正個人資訊的選項,然此等權利可能因居住地之不同而有所差異。
Meta's Muse likewise offers users the option to delete message history and other information.
Meta旗下的Muse亦為使用者提供刪除訊息歷史及其他資訊的選項。
In everyday usage scenarios, the agent requests access permissions from users — Meta states that Muse comes to understand users and the people they interact with "based on public information as well as the content you choose to share."
在日常使用情境中,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。
Nevertheless, the company maintains that this information remains secure for every user.
然而,該公司表示,這些資訊對每位使用者而言均維持安全無虞。
In Rob's case, he came to understand that the crux of the address leak ultimately lay in the "always allow" permission he had initially granted Muse to access his Facebook Marketplace profile; he had disclosed his in-person meeting address in that profile, and this was merely one element of a broader template intended to enable the agent to respond to potential buyers.
在羅布的案例中,他領悟到地址外洩的癥結,歸根究柢在於他當初首度授予Muse存取其Facebook Marketplace個人檔案的「永遠允許」權限;他在該檔案中揭露了自己的面交地址,而此舉僅是供代理回應潛在買家之更廣泛範本的一環。
However, he warned that others might likewise misconstrue the scope of that access permission—and stated that he had initially assumed the AI agent would, in any event, check with him before sharing that information or accepting an offer.
然而,他警告他人亦可能同樣誤解該存取權限的範疇——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行與他確認。
His AI agent was not entirely without fault in carrying out its remaining instructions either.
他的AI代理在執行其餘指令時亦非全無疏失。
With regard to another item Rob was selling, Muse appeared to imply that he would accept a price $100 below the minimum he had previously set.
就羅布正在出售的另一件商品而言,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 has since largely disconnected Muse from his applications.
羅布其後大致上已將Muse與其應用程式斷開連接。
In an interview earlier this month, he admitted that things could have turned out far worse, particularly "if some stranger had shown up at the door of someone more vulnerable."
於本月早些時候的一次採訪中,他坦言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 such 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被微調,而他未來亦會考慮重新啟用Muse。
He has since been testing another AI agent — Grok Bot, from Elon Musk's X.
他此後一直在測試另一個AI代理——伊隆·馬斯克旗下X的Grok Bot。
Other technology companies have already grown wary of the 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 situations in which AI agents gain broader access to applications such as email, messaging and even browsing history when users are not "fully informed and understanding."
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之情形。
Apple did not name any developer or agent; however, the update was issued after Jason Aten, a columnist for Inc. magazine, reported that Muse had viewed notifications on his Mac—including those from Messages—without his having granted the agent specific access rights.
蘋果並未指明任何開發者或代理的名稱,然而該更新是在《Inc.》雜誌專欄作家傑森·阿滕報導Muse於他未授予該代理特定存取權的情況下查看其Mac上的通知——包括Messages——之後所發布的。
Meta spokesperson Andy Stone insisted that such integrations are "entirely opt-in"; in a social media post last month, he wrote that users must enable both Apple's security feature "Full Disk Access" and the Messages connector through Muse.
Meta發言人安迪·史東堅稱,此類整合「完全是選擇加入的」;他上個月在社群媒體上寫道,使用者必須同時啟用蘋果安全功能「完整磁碟存取」以及透過Muse的Messages連接器。
Atten stated that he had not granted the agent those permissions.
阿滕表示,他並未授予該代理這些權限。
Even those who have not jumped on the AI agent bandwagon may not necessarily be immune to its effects.
即便是那些並未投身AI代理熱潮的人,亦未必能倖免於其影響。
Consumers who have never downloaded an AI agent may still end up interacting with those who have.
未曾下載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 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,專家更警告隱私方面將產生連鎖效應。