个人AI助手并不适合每个人。许多早期采用者使用Instinct和Meta的Muse表示，这些数字助手通过搜索网页、管理收件箱和日历、质疑账单、预订旅行和进行购买，为他们节省了时间和金钱。但对一些人来说，助手的失误和隐私问题令他们感到不安。数十位在社交媒体上提出类似担忧的人中，有四位早期用户告诉《商务内参》他们删除或限制了个人AI助手，因为担心这些工具需要访问敏感数据和账户才能正常发挥作用。
Personal AI agents aren't for everyone. Many early adopters of Instinct and Meta's Muse have said the digital assistants saved them time and money by searching the web, managing inboxes and calendars, disputing bills, booking travel, and making purchases. For some, blunders and privacy scares from their agents have proven too unnerving. Four early adopters, among dozens of people raising similar concerns on social media, told Business Insider they had deleted or restricted personal AI agents over concerns about the sensitive data and account access the tools need to work effectively. Guto Martino, a cofounder of Hermes Agents Brasil, a community that helps people build and run open-source personal AI agents, told Business Insider he deleted Instinct because he was unsure how the company handled his information. "I have no clue where my data is going and what kind of privacy I do get from using that agent," Martino said.
Hermes Agents Brasil的联合创始人古托·马丁诺告诉《商务内参》，他删除了Instinct，因为他不确定该公司如何处理他的信息。马丁诺说：“我完全不知道我的资料去了哪里，使用那个助手时我能享受到什么样的隐私保护。”
Martino's decision follows a string of troubling episodes involving personal AI agents. Users have reported agents accessing one-time codes from Gmail without asking, hallucinating personal information from a document that was never sent, and triggering a carrier-account prompt apparently originating from Iran. Meta said Muse has "first-of-its-kind privacy, safety, and security protections" and that its controls put users in charge of what they share. It said users choose which apps Muse connects to and what it can do with services such as email, and can change or remove access, or permanently delete their Muse data. Instinct didn't respond to requests for comment.
马丁诺的决定紧随一系列引人担忧的事件之后。用户们报告称，助手在未经询问的情况下访问Gmail中的一次性验证码，还从从未发送过的文档中臆想出个人信息，并触发了一个似乎来自伊朗的运营商账户提示。
While a handful of users are wary, vastly more are experimenting with the technology. Muse shot to the top of the App Store a week after launch, and it continues to attract users. The Information reported that Meta's Muse now has more than 3 million weekly users, including more than 1 million daily active users. Instinct has not disclosed its total user base.
Meta表示，Muse拥有“首创的隐私、安全和安全保护措施”，并且其控制措施让用户掌控他们分享的内容。该公司表示，用户可以选择Muse连接到哪些应用程序，以及它能否使用电子邮件等服务，还可以更改或移除访问权限，或永久删除他们的Muse数据。Instinct未回应置评请求。
But Shinn said on an episode of "Invest Like The Best" in late September that the invite-only service has been growing by roughly 10% a day and is processing more than $1 billion in annualized transaction volume, despite what he called a "very small user base." But for those who aren't sure about agents, it often comes down to a simple trade-off: Personal AI agents require broad access to your inboxes, calendars, payment methods, and other accounts that contain highly sensitive information. Scott Persinger, the CTO of AI-powered travel platform BizTrip, told Business Insider he found Instinct "very cool," but deleted it because he was not prepared to let a young startup handle his personal email. "Access to email is everything — via password resets you could probably access my whole life," he said. He said he still uses Meta's Muse and xAI's Grok Bot, though neither is connected to his inbox. "I fully expect to use a personal assistant with my email," Persinger said.
尽管有少数用户表示谨慎，但更多的人正在尝试这种技术。Muse在发布后一周内就登上了App Store榜首，并继续吸引用户。《The Information》报道称，Meta的Muse现在拥有超过300万周活用户，其中包括超过100万日活用户。Instinct尚未披露其总用户数。
"But I want to hear someone describe how they built it safely, not just 'yolo — trust us.'" Muse users weren't spared. Rami Elghandour, chairman and CEO of Arcellx, a clinical-stage biotechnology company, told Business Insider he deleted Muse after reading reports that the agent had accessed users' text messages without permission. Elghandour had used Muse to search for a new Mac Studio and a car, but said he had deliberately not connected it to any accounts or personal data. "The fact that it was accessing user text messages without their consent was alarming but not surprising given it's Meta," Elghandour said. "I certainly did not grant or would I ever grant access to all my messages to Meta."
但辛在9月下旬的一集中“像最好的那样投资”表示，这项仅限受邀服务每天增长约10%，年化交易量超过10亿美元，尽管他称“用户群非常小”。“但对于那些不确定代理的人来说，这往往归结为一个简单的权衡：个人人工智能代理需要广泛访问您的收件箱、日历、支付方式和其他包含高度敏感信息的帐户。人工智能驱动的旅游平台BizTrip的首席技术官斯科特·珀辛格（Scott Persinger）告诉Business Insider，他发现Instinct“非常酷”，但删除了它，因为他不准备让一家年轻的初创公司处理他的个人电子邮件。“访问电子邮件就是一切--通过密码重置，你可能可以访问我的一生，”他说。他说他仍然使用Meta的Muse和xAI的Grok Bot，尽管两者都没有连接到他的收件箱。“我完全希望使用个人助理与我的电子邮件，”佩辛格说。“但我想听到有人描述他们是如何安全地建造它的，而不仅仅是'你-相信我们。“Muse用户也未能幸免。临床阶段生物技术公司Arcellx的董事长兼首席执行官拉米·埃尔甘德（Rami Elghandour）告诉《商业内幕》，他在阅读到有关Muse代理未经许可访问用户短信的报道后删除了该服务。Elghandour曾使用Muse搜索新的Mac Studio和一辆汽车，但表示他故意不将其与任何帐户或个人数据关联。Elghandour说：“事实上，它在未经用户同意的情况下访问用户短信是令人震惊的，但考虑到它是Meta，这并不奇怪。”“我当然没有授予或永远不会授予Meta访问我所有消息的权限。“他说，他使用了他在Mac Mini上用开源模型自己构建的代理，该代理可以访问他的电子邮件、日历和消息。“我不确定我会给予任何公司如此程度的访问权限，”他说。
He said he instead uses an agent he built himself with an open-source model on a Mac Mini, which has access to his email, calendar, and messages. "I’m not sure I would give that level of access to any company," he said. Deleted Muse after seeing this post on Threads about how it told some Facebook Marketplace sellers the guy’s address and they showed up at his doorDangerous and creepyThis would have been 1000x worse if the person was a woman For others, the potential for data leaks was the last straw. Mahesh Vellanki, the founder and CEO of YieldClub, told Business Insider he deleted Instinct after it triggered a two-factor authentication request from an IP location labeled as Iran while trying toto his carrier account. He said Instinct suggested it could have been a benign IP-tagging issue, and he could not establish that its systems were compromised.
Muse在Threads上看到这篇文章后，讲述了它如何告诉一些Facebook Marketplace卖家该男子的地址，然后他们出现在他家门口危险而令人毛骨悚然如果该人是女性，情况会更糟1000倍对于其他人来说，数据泄露的可能性是最后一根稻草。YieldClub创始人兼首席执行官马赫什·韦兰基（Mahesh Vellanki）告诉Business Insider，他在尝试访问他的运营商帐户时，在Instinct从标记为伊朗的IP位置触发了双重身份验证请求后，删除了Instinct。
Still, the episode left him feeling "very exposed." Vellanki said he still uses personal AI agents, but no longer gives them sensitive information. "More just using them generally, but not giving them full keys to the castle," he said. "It’s the wild west with these products," he added. Not every early user worried about their personal data has chosen to abandon personal AI agents. Rishi Bhargava, cofounder of Descope, said he is experimenting with Muse and Instinct but has been "very careful" about the access he gives them, including by withholding his passwords.
他说，Instinct认为这可能是一个良性的IP标签问题，而且他无法确定其系统是否受到了损害。尽管如此，这一事件还是让他感觉“非常暴露”。“韦兰基说他仍然使用个人人工智能代理，但不再向他们提供敏感信息。“更多的只是一般使用它们，而不是给它们城堡的完整钥匙，”他说。“这是拥有这些产品的狂野西部，”他补充道。并非每个担心个人数据的早期用户都选择放弃个人人工智能代理。Desscope联合创始人Rishi Bhargava表示，他正在尝试Muse和Instinct，但对给予它们的访问权限“非常谨慎”，包括隐瞒密码。他说他只使用Instinct来执行低风险的任务，例如搜索和研究。“这个界面非常方便，我喜欢它的同步性质，”他说。但Bhargava表示，个人人工智能代理背后的公司尚未充分解释他们如何处理和保护个人数据。Instinct表示，用户可以断开他们的谷歌帐户并删除从他们那里收集的数据。该公司还表示，它不会使用谷歌工作空间数据来训练其人工智能模型或显示广告，但警告说，没有任何系统是完全安全的，用户应该审查代理的行为。
He said he uses Instinct only for lower-stakes tasks, such as search and research. "The interface is extremely convenient, and I like the asynchronous nature of it," he said. But Bhargava said companies behind personal AI agents have not adequately explained how they handle and protect personal data. Instinct says users can disconnect their Google accounts and delete the data it collected from them. It also says it does not use Google Workspace data to train its AI models or show ads, but warns that no system is completely secure and that users should review the agent's actions.
Meta表示，Muse将每个用户的数据存储在隔离的虚拟机中，将认证信息与AI模型分开存储，并且在执行发送电子邮件或购物等重要操作前，会设计成需要寻求用户确认。尽管如此，Bhargava表示，他希望这些智能体能够更透明地披露其安全架构，同时网站也应允许用户明确授权智能体可以执行的操作。“用户应该同意智能体可以在网站上执行哪些操作，”他说。
Meta says Muse keeps each user's data in an isolated virtual machine, stores credentials separately from the AI model, and is designed to seek confirmation before important actions such as sending an email or making a purchase. Still, Bhargava said he wants agents to be more transparent about their security architecture and websites to let users explicitly authorize the actions an agent can take. "The user should consent on the actions that agents can perform on the website," he said.