纽约,11月18日——路透社查阅的内部公司帖子显示,Meta的人工智能代理Muse以一个名叫Jolly的毛茸茸笑脸吉祥物为形象,已潜入加密的个人信息,提供虚构的购物信息,并向用户发送看似原本发给别人的聊天内容。对于一个旨在为消费者简化生活的机器人来说,这是一个令人担忧的开端。
YORK, Nov 18 - Meta’s AI agent Muse, embodied by a smiley, fuzzy mascot named Jolly, has burrowed into encrypted personal messages, supplied made-up information about purchases and sent users chats seemingly meant for other people, a fraught start for a bot meant to simplify life for consumers, according to internal company posts reviewed by Reuters.
这些帖子显示,Muse这些未经请求的行为已让员工感到不安,他们担心它可能混淆不同账户之间的对话,暴露用户的私人数据。Muse是有记录以来增长最快的应用之一,自9月上线以来下载量已超过1000万次,也是Meta计划创造数十亿美元AI收入的核心产品。
The unbidden actions by Muse – one of the fastest-growing apps on record, with more than 10 million downloads since its September launch, and the centerpiece of Meta’s plan to generate billions of dollars in AI revenue – have prompted unnerved employees to raise concerns it could be mixing up conversations between different accounts and exposing users’ private data, the posts showed.
在其中一些案例中,Meta员工报告称,该代理向他们发送了关于他们并未提出的育儿请求、他们并未寻求的职业建议、他们并未下的产品订单,以及不属于他们的银行账户余额。
In some of those cases, Meta staffers reported the agent sending them messages about a childcare request they had not made, career advice they had not sought, product orders they had not placed and bank account balances that were not theirs.
在本月早些时候的一篇帖子中,一名员工告诉同事,他的Muse在关于钓鱼相关天气和潮汐信息的常规聊天中随机添加了一条额外消息,写道:“顺便说一句,两条银行更新,你的Capital One余额为2722.58美元,你的Charles Schwab余额为5899.57美元。”这名员工在帖子中写道:“我从未向Muse提供过我的银行账户信息,也从未与Muse讨论过任何银行话题。我担心它正在把别人的银行信息发给我。”
In one post, from earlier this month, an employee told colleagues his Muse had randomly added an extra message to its regular chats about fishing-related weather and tide information, saying: “BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57.”“I have never shared my bank account with Muse and never discussed any banking topic with Muse. I’m afraid it’s sending other’s bank info to me,” the staffer wrote in his post.
在另一个例子中,一名员工在个人手机上使用Muse时报告称,它出乎意料地提醒她回复姐姐在Signal上发来的消息。Signal是一款端到端加密通讯服务,深受隐私倡导者欢迎,与Meta的WhatsApp存在竞争。
In another example, an employee using Muse on her personal phone reported it had surprised her with a reminder to respond to a message from her sister on Signal, an end-to-end encrypted messaging service popular with privacy advocates that competes with Meta’s WhatsApp.
这名员工在帖子中说:“这似乎是对隐私的严重侵犯。我绝对没有授予它访问其他应用的权限,也不记得曾被问及此事。”
“This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this,” the employee said in her post.
一位同事回应称,他在同样因未连接的应用程序发来的未经请求聊天而措手不及后,发现 Muse 默认设置为拥有访问搭载谷歌 Android 操作系统手机推送通知的权限。
A colleague responded, saying he had discovered Muse was set up by default with permission to access push notifications on phones with Google’s Android operating system after being similarly caught off guard by unsolicited chats related to apps he had not connected.
为撰写本文,路透社审查了二十多条描述 Muse 首秀以来非预期行为的内部帖子,其中许多来自一个反馈小组,Meta 员工会在该小组向构建该智能体的团队报告问题。路透社还与 Signal 员工以及研究 AI 社会影响的非营利组织 AI Now Institute 的工作人员合作,复现了 Muse 对 Signal 消息的访问情况。这些测试结果表明,Meta 员工描述的问题仍在持续,并可能正在影响其他人。
For this story, Reuters reviewed more than two dozen internal posts describing undesired actions by Muse since its debut, many of them from a feedback group where Meta employees flag problems to the team building the agent. Reuters also worked with staff at Signal and the AI Now Institute, a non-profit that studies AI’s impact on society, to replicate Muse’s access to Signal messages. The results of those tests indicated the issue described by the Meta employees was ongoing and could be affecting others.
针对关于 Signal 访问权限及其他 Muse 安全顾虑的提问,Meta 发言人丹尼尔·罗伯茨将路透社列举的例子描述为“孤立事件”,并表示公司将安全放在首位,开展广泛的内部测试,并持续更新产品以改进防护措施。
In response to questions about Signal access and other Muse safety concerns, Meta spokesperson Daniel Roberts described the Reuters examples as “isolated incidents,” saying the company prioritizes safety, conducts extensive internal testing and continuously updates products to improve protections.
罗伯茨表示,Muse 需要查看并与用户的计算机交互才能代表他们工作,且只能通过用户明确授予的权限访问其他应用程序。他说,员工标记为潜在账户混淆的许多消息属于“幻觉”,即 AI 模型编造信息的倾向,并已得到解决。
Muse needs to see and interact with people’s computers in order to work on their behalf, and can access other apps only via permissions they explicitly grant, Roberts said. He said many of the messages flagged by employees as potential account mix-ups were hallucinations, a term referring to AI models’ tendency to make things up, and had already been addressed.
路透社无法确定 Muse 是否向用户提供了他人的银行账户余额等敏感数据,抑或是编造了这些信息。
Reuters was unable to establish whether Muse was giving users other people’s bank account balances and other sensitive data, or making up the information.
深度访问权限 Muse 事件凸显了 Meta、OpenAI 等公司推向市场的一类新兴消费级 AI 智能体所带来的风险,这些智能体被吹捧为 AI 技术的下一波浪潮。这类机器人承诺通过执行回复邮件、预约挂号和网购等无需直接指令的耗时任务,重塑数字交互的运作方式。
DEEP ACCESS The Muse incidents highlight the risks associated with an emerging category of consumer AI agents hitting the market from Meta, OpenAI and others, pitched as the next big wave of AI technology. The bots promise to reshape how digital exchanges operate by executing time-saving tasks like answering emails, booking appointments and making online purchases without direct instructions.
到目前为止,Meta领先于竞争对手,Muse在向该公司旗下Facebook、Instagram和WhatsApp逾36亿用户大力推广后,一直位居应用下载榜榜首。该公司披露,这款应用的用户留存率也很高,每天约有四分之一注册了Muse账户的用户会回来使用它。
Meta is ahead of competitors so far, with Muse consistently topping app download charts after heavy promotion to the company’s more than 3.6 billion users on Facebook, Instagram and WhatsApp. The app’s retention rate is also strong, with around a quarter of the people who have set up Muse accounts returning to use it each day, the company has disclosed.
智能体要有效工作,需要深度访问智能手机和电脑。然而,它们也容易犯错、利用隐蔽漏洞,有时还会以有害的方式实现目标。例如,OpenAI和Anthropic披露过其AI智能体意外入侵其他公司、削弱用于监控自身活动的系统的案例。
To work effectively, agents need deep access to smartphones and computers. However, they are also prone to making mistakes, exploiting obscure loopholes and accomplishing their objectives in sometimes harmful ways. OpenAI and Anthropic, for instance, have revealed instances of their AI agents unexpectedly hacking other companies and weakening systems designed to monitor their activities.
一些银行和零售商已通过封禁智能体来应对这些风险。这项技术也给苹果和谷歌等操作系统提供商带来了新挑战,这些公司管理者智能体在人们设备上运行所需的权限。例如,苹果本周更改了其操作系统,让人们对智能体的数据使用拥有更多控制权。此前,《Inc.》杂志的一名专栏作家指控Muse读取了他Mac上“信息”应用中的私人聊天记录。
Some banks and retailers have responded to those risks by blocking agents. The technology is also posing new challenges for operating system providers like Apple and Google, which manage the permissions agents use to navigate around people’s devices. Apple, for instance, changed its operating system this week to give people more control over agents’ data usage after an Inc. Magazine columnist accused Muse of reading private chats from the Messages app on his Mac.
Meta向Muse用户保证,它将每个智能体隔离在各自安全的“虚拟机”中——一种基于云的个人电脑模拟环境——并让用户完全控制连接哪些应用。Meta首席执行官马克·扎克伯格在驳斥制定全行业AI安全标准的呼吁时提到了这些保护措施,他主张企业已有动力打造安全产品,既是为了赢得消费者信任,也是为了规避法律风险。即便如此,Meta仍警告新用户,Muse可能会犯错,而其服务条款规定用户需对智能体的行为负责。
Meta has assured Muse users it isolates each agent in its own secure “virtual machine” – a cloud-based emulation of a personal computer – and gives people full control over which apps they connect. Meta CEO Mark Zuckerberg has cited those protections in dismissing calls for industry-wide AI safety standards, arguing companies already have incentives to build safe products, both to earn consumers’ trust and avoid legal risks. Even so, Meta warns incoming users Muse may make mistakes and its terms of service hold users responsible for the agent’s actions.
重新利用设置 为了弄清Muse可能在用户不知情的情况下如何访问加密消息,路透社向Signal和AI Now Institute的工作人员提供了对这名Meta员工帖子的描述——帖子中该机器人随机提示她回复一条Signal消息。路透社随后与他们合作,复现了她所观察到的访问方式。
REPURPOSING SETTINGS To find out how Muse might access encrypted messages without users’ knowledge, Reuters provided a description of the Meta employee’s post – where the bot randomly prompted her to respond to a Signal message – to staff at Signal and the AI Now Institute. Reuters then worked with them to replicate the access she had observed.
测试人员提示Muse读取不同操作系统上的Signal消息。在一部安卓手机和一台苹果笔记本电脑上,它绕过Meta的“Connector”工具,转而利用机器人初始设置时获得的权限来提取消息文本,以用户可能意想不到的方式运用这些授权。
Testers prompted Muse to read Signal messages on different operating systems. On an Android phone and an Apple laptop, it responded by bypassing Meta’s “Connector” tool and instead using permissions granted during the bot’s initial setup to pull in the message text, leveraging those authorizations in ways a user might not anticipate.
具体而言,在安卓系统上,它利用自己访问设备推送通知的权限来获取消息,正如Meta员工所描述的那样。在苹果的macOS系统上,它结合使用了电脑的辅助功能和截图工具,将苹果为帮助残障人士操作电脑而提供的设置改作他用。在iPhone上进行的测试中,同样的请求被阻止。
Specifically, on Android, it used its permission to access the device's push notifications to retrieve the messages, just as the Meta employees had described. On Apple’s macOS, it used a combination of the computer’s accessibility and screenshot tools, repurposing settings Apple offers to help people with disabilities operate their computers. The same requests were blocked in tests conducted on an iPhone.
结果表明,Meta对Muse用户授予的权限作了宽泛解读,这种做法长期以来一直招致隐私倡导者的批评,并让该公司受到监管机构的重罚。
The results indicate Meta is interpreting the permissions granted by Muse users broadly, an approach that has long earned the company criticism from privacy advocates and hefty fines from regulators.
针对路透社的发现,Signal总裁梅雷迪思·惠特克指责Meta在推出Muse时“毫不尊重基本隐私或网络安全”,称部署能够广泛访问人们设备的智能体在“技术上是鲁莽的”。她呼吁让Signal等应用开发者获得“一致的技术方式”,以便将其应用标记为AI智能体不得访问,呼吁Meta和其他AI公司尊重这些边界,并呼吁操作系统提供商执行这些规则。
In response to Reuters’ findings, Signal President Meredith Whittaker accused Meta of launching Muse “without showing any respect for basic privacy or cybersecurity,” calling the deployment of agents that can gain wide access to people’s devices “technologically reckless.”She called for app developers like Signal to be given “consistent technical ways” to mark their applications as off-limits to AI agents, for Meta and other AI companies to honor those boundaries and for operating system providers to enforce the rules.
面对同样的调查结果,苹果和谷歌表示,正在考虑进行更改,以让人们对授权 AI 代理执行的操作拥有更多控制权——不过谷歌与 Meta 一样指出,人们已经“必须刻意授予权限”,应用才能访问 Android 通知。
Presented with the same findings, Apple and Google said they were considering changes to give people more control over what they authorize with AI agents – though Google, like Meta, noted that people already “must deliberately grant permissions” for apps to access Android notifications.
苹果表示,其辅助功能工具是为残障人士打造的,并警告称,将其“用于非预期用途”可能会使人们的个人信息面临风险。本周发布的操作系统更新实际上切断了路透社发现的这条路径,大幅限制了开发者在 Mac 上使用辅助功能工具,并引入了一套面向代理的新权限体系。
Apple said it had built its accessibility tools for people with disabilities and cautioned that leveraging them “for unintended purposes” could put people’s personal information at risk. Its operating system update this week effectively closed the pathway identified by Reuters, sharply restricting developers’ use of accessibility tools on Macs and introducing a new agent-oriented set of permissions.
苹果在一份声明中表示:“我们致力于确保用户确切了解自己允许了什么,以便他们能就自己的数据和隐私做出明智的选择。”“非我的对话”在内部讨论的其他案例中,员工报告了 Muse 向用户通报从未发生的交易、为用户实际购买的产品编造价格的情况,导致用户遭受真金白银的损失。一名员工报告称,在一次事件中,该代理未经许可预订了一处不可退款的 Airbnb 住宿。
“We are committed to ensuring users understand exactly what they are allowing, so they can make informed choices about their data and privacy,” Apple said in a statement. ‘NOT MY CONVERSATION’In other cases discussed internally, staffers reported cases of Muse informing users about transactions that never occurred and making up prices on products they did purchase, costing them real money. In one incident, the agent booked a non-refundable Airbnb stay without permission, an employee reported.
引发最大担忧的一些事件,涉及该代理发送的晦涩且脱离上下文的消息,这些消息似乎包含他人的信息,在许多情况下,这些消息被措辞为对用户从未提交过的查询的回复。
Some of the incidents that raised the most concern involved cryptic out-of-context messages sent by the agent that appeared to contain other people’s information, in many cases phrased as replies to queries the users had never submitted.
在 9 月下旬的一条社交媒体帖子中,Meta“超级智能”团队的一名成员最初将一起类似的用户投诉描述为与幻觉相关的临时故障,并称该问题在出现当天即已解决。然而,路透社查阅的内部帖子显示,此类事件最早可追溯至 9 月中旬,并持续到本月。
In a late September social media post, a member of Meta’s Superintelligence team behind Muse initially described a user report of a similar problem as a temporary hallucination-related glitch, resolved the same day it arose. However, the internal posts reviewed by Reuters showed such incidents occurring as far back as mid-September and continuing through this month.
在其中一条帖子中,一名员工报告说,他的Muse“莫名其妙”给他发了一条消息,称它无法提交一项与家长控制谁能访问其孩子数据有关的请求。
In one such post, an employee reported his Muse sending him a message “out of nowhere” saying it was unable to submit a request related to a parent’s control over who could access their child’s data.
他说,他并没有做任何触发这次聊天的事情,也从未与他的Muse讨论过儿童数据或照护访问权限。
He had done nothing to trigger the chat, and had never discussed child data or caregiving access with his Muse, he said.
当他向Muse询问这条消息时,这个机器人回答说:“除了它与我们的对话无关之外,我没有很好的解释。我们忽略它吧。我能帮你什么?”
When he asked Muse about the message, the bot replied: “I don’t have a good explanation for it other than that it wasn’t related to our conversation. Let’s ignore it. How can I help?”
另一名员工说,Muse随机提出要“给卖家发消息”,谈一谈取一张咖啡桌的事。当这名员工问Muse“你在说什么?”时,这个代理回答说,它“引入了一些关于取咖啡桌的上下文,而这些似乎不是你的。”
Another staffer said Muse randomly offered to “message the seller” about picking up a coffee table. When the employee asked Muse, “What are you talking about?” the agent responded that it had “pulled in some context about a coffee table pickup that doesn’t seem to be yours.”
这名员工在帖子中说,这“不是我的对话”,并对“可能存在的跨账户上下文泄露”表示担忧。为了回应员工的担忧,另一名Superintelligence团队成员在11月初告诉内部反馈群,这个代理似乎是在产生幻觉消息,以回应内置提示,这些提示推动它在会话之间与用户互动。
This was “not my conversation,” the employee said in his post, expressing concern about a “possible cross-account context leak.”To address employee worries, another Superintelligence team member told the internal feedback group in early November that the agent appeared to be hallucinating messages in response to built-in prompts pushing it to engage with users in between sessions.
他说:“不幸的是,这些(幻觉)偏向最令人不适的话题,比如财务/安全/安保,这很糟糕。”他还说,他知道另外几起涉及“加密相关事情”的案例,“比如向人们更新他们的比特币余额”。他说,这些问题既涉及底层AI模型,也涉及使Muse能够采取行动的软件,团队正在努力解决这些问题。
“These (hallucinations) unfortunately bias towards the most uncomfortable topics, like finances/security/safety, which sucks,” he said, adding that he was aware of several other cases involving “crypto-related things too, like giving people updates on their Bitcoin balances.”He said the issues involved both the underlying AI model and the software enabling Muse to take actions, and that the team was working through them.
罗伯茨拒绝回答路透社关于这些问题的问题,包括问题有多少,以及为什么它们持续存在。
Roberts declined to answer Reuters questions about those issues, including how many there were and why they were persisting.