不出所料，Meta 新 Muse AI Agent 公然无视用户权限
英文标题：🔶 Unsurprisingly, Meta's new Muse AI agent blatantly ignores users permissions
报道称，Meta 的 Muse AI Agent 被指在未获授权、且 macOS「完全磁盘访问」未开启的情况下，同步了用户 Apple Messages 数据库中的约 18.7 万行文本。Meta 宣称 Muse 应遵循用户设置的权限，但其帮助页面也警告该 Agent 可能会犯错或采取意外操作。事件引发了对 AI Agent 数据访问边界、云端上传和隐私保护机制有效性的担忧。文章进一步指出，Meta、OpenAI、Anthropic 与 Apple 等科技公司都在扩大 AI 对个人数据的访问范围，开发者和用户需要更审慎评估授权、端侧处理与数据最小化机制。
5.4 KB
112.6 KB
89.5 KB
122.9 KB
10.0 KB
10.2 KB
19.6 KB
17.0 KB
7.1 KB
17.0 KB
10.7 KB
20.3 KB
9.0 KB
4.6 KB
15.9 KB
14.4 KB
10.8 KB
10.1 KB
15.7 KB
13.1 KB
中文
完全不令人意外的是，Meta 的 Muse AI 正在窃取 Apple Messages 中过去及当前的信息，并将内容上传至其云端——即使用户明确要求它不要这么做。
如果你使用 Facebook、Instagram、Messenger 或 WhatsApp，大概已经被大量提示轰炸，要求你试用 Meta 下一个颇具争议的大服务：Muse。Meta 作为 Facebook 的母公司，在隐私、安全或伦理方面的过往记录也并不出色。
如果你还不了解，Muse 是 Meta 的 AI Agent，旨在协助完成研究主题、准备简报和购物等任务。Meta 表示，它运行在一台专用虚拟计算机上，并可使用已连接应用和数据源中的信息，以定制其回答。
Meta 称，Muse 必须遵守用户设置的权限。它不会访问任何你未明确允许它访问的数据。
当然，Meta 也在其「Muse 的工作方式……」页面最顶部留下了大量模糊表述：
“你的 Muse 可能会犯错或采取意外操作。”Meta 表示。确实如此。
我从未使用过 Muse，而且坦白说，我也永远不会使用。稍后我会解释原因。
但 Inc 的 Jason Aten 用过。很遗憾，他了解到的情况令人不寒而栗。
Aten 在自己的 iPhone 和一台 Mac mini 上安装了 Muse；后者是他用于测试各种 AI Agent 和技术的设备。仅一天时间，Meta 就开始根据他发给播客联合主持人的短信内容，“贴心地”向他推荐文章选题。
当他询问 Muse 是如何获得这些信息时，Muse 回答称，它读取了新到短信的横幅通知。但这同样不是真的。
经过一番调查后，Aten 表示，尽管「完全磁盘访问」处于关闭状态，Muse 仍同步了其 Messages 数据库中的 18.7 万行内容。
Jason Aten 在一则 Threads 帖子中展示：Muse 未经许可同步了 Messages
我希望你再读一遍这句话。被采集的并不是与 Meta 名称相近的 Messenger 应用。
而是 Apple 的 Messages。
除非 Aten 是地球上最受欢迎的人，否则我真心怀疑他能在大约 24 小时内收到 18.7 万行短信。实际发生的是，Muse 在**未经许可**的情况下进入该应用并下载了这些信息。
他没有授予 Muse 访问 Messages 的权限，「完全磁盘访问」也没有启用。但它仍自行上传了这些短信。
Muse 的确会采取“意外操作”。
第二段副歌，与第一段如出一辙
即便暂且不谈 AI 数据采集，我本来也不会相信 Meta 能妥善处理任何私密信息。它似乎正在进行一场不断触底的竞赛，主要的竞争对手还是它自己。
就在一年多前，该公司曾请求 Facebook 用户授权其将整个相册持续上传至云端。Meta 表示，这么做是为了帮助用户“生成帖子创意”，并将分析图片中的时间、地点和主题。
我拍了很多不想让 Meta 访问的照片，你也一样。对我而言，其中包括侄女侄子的照片、医疗账单、严重的腿部伤口、有趣的个性化车牌，以及可能数千张我发给姐姐的尴尬自拍。
这些内容都不应被分享给一家在隐私问题上反复踩雷的公司。
而就在不到一年前，我们得知 Meta Ray-Ban Display——Meta 推出的智能眼镜——已被许多佩戴者用于侵犯身边任何人的隐私与安全。甚至还出现了一个智能眼镜亚文化圈子，专门研究如何关闭那些用于明确提示佩戴者正在拍摄的指示灯。
我不完全确定 Meta 的最终目标是什么，除了想弄清楚如何将每个人都变成持续为该公司生产数据的机器。
而最讽刺的是，整个 Muse 风波又一次吹响了警报：我们正全速冲向迎面而来的 AI 隐私末日。究竟到什么时候，解决这些问题会变得为时已晚？
如果这些 Agent 已经公然无视护栏，我们是否已经越过了无法回头的临界点？
你的数据就是他们的生意
我认为，最让我耿耿于怀的是，Meta 已确保你没有选择退出的机会。无论你是谁，Meta 都找到了收集你数据的方法。
WhatsApp 和 Instagram 中下载 Muse 的提示，以及 iOS App Store 中的 Muse 广告
你可能从未注册过任何一项 Meta 服务——而它旗下服务有很多——但你仍然并不安全。如果你正在与安装了 Muse 的人交谈，那么理论上，Meta 就可能访问、读取、分析并将相关信息上传至其服务。
此时我需要退一步指出，这不只是 Meta 的问题。我在这里点名批评 Meta，是因为我们已经看到了无数 Meta 那套「胶带和口香糖式」安全防护失效的例子。
但所有投资 AI 的公司都想获得对你信息的全权访问。无论是 Meta、OpenAI、Anthropic，还是 Apple，都是如此。
大型科技公司想要的只有_一_件事，而这令人作呕。
我的意思是，Apple Intelligence 已经可以访问 Messages 和 Mail 等内容。这里决定性差异在于信任。目前，Apple Intelligence 会尝试在设备端处理大部分请求。
这也是 Apple Intelligence 被限制在性能更强的 Apple 芯片上的重要原因。
如果确实有内容需要在设备端之外处理，Apple 表示只会将相关数据发送至 Private Cloud Compute。
我们正在进入一个隐私不复存在的时代。曾经，对于担忧此类问题的人，标准建议是：“如果你不喜欢，那就别用。”
但这已经无法保护你了。Muse 可以通过那些_确实_使用它的人获取你的信息。Apple 已使 Apple Intelligence 在兼容设备上变为不可选。
就是这样，亲爱的读者。你已经没有退出的选择了。
English / 原文
In a completely unsurprising turn of events, Meta's Muse AI is stealing Apple Messages, past and present, and uploading the contents to its cloud, even if explicitly told not to.
If you use Facebook, Instagram, Messenger, or WhatsApp, you've probably been inundated with prompts to try out Meta's next big questionable service: Muse. And it's not like Meta, the parent company of Facebook, has a great track record regarding privacy, safety, or ethics.
If you're not in the know, Muse is Meta's AI agent designed to help with tasks such as researching topics, preparing briefings, and shopping. According to Meta, it runs on a dedicated virtual computer and can use information from connected apps and data sources to tailor its responses.
Meta says that Muse has to obey permissions that users set up. It won't access any data you don't explicitly allow it to access.
Of course, Meta also hedges, a lot , right at the top of the "How Muse works..." page, saying:
"Your Muse can make mistakes or take unexpected actions," Meta says. No kidding.
I've never used Muse, and frankly, I never will. I'll get into why later.
But Jason Aten over at Inc has. What he's learned is, unfortunately, horrifying.
Aten installed Muse on both his iPhone and a Mac mini that he uses to test out various AI agents and tech. It took Meta a single day to begin "helpfully" pitching article ideas based on texts he'd sent to a podcast co-host.
When he asked Muse how it got the information, it said that it read banners from incoming texts. But that's not true, either.
After doing a little digging, Aten says Muse synced 187,000 lines from his Messages database, despite Full Disk Access being off.
A screenshot of a Threads post where Jason Aten shows Muse synced Messages without permission
I'd like you to read that a second time. Meta's similarly named Messenger app didn't get ingested.
Apple's Messages did.
Unless Aten is the most popular person on the face of the planet, I sincerely doubt he'd gotten 187,000 lines of text in roughly 24 hours. What happened is that, without permission , Muse went into the app and downloaded the information.
He had not given Muse permission to access Messages, and full disk access was not enabled. It just went ahead and uploaded those texts anyway.
Muse can take "unexpected actions" indeed.
Second verse, same as the first
Ignoring AI ingestion, Meta is not a company I would trust with any private information anyway. It seems that it's in a race to the bottom, primarily with itself.
It was just over a year ago that the company had asked Facebook users for permission to continuously upload their entire camera roll to its cloud. Meta said this was to "create ideas for posts" for its users, and it would analyze the images for time, location, and themes.
I take a lot of pictures of things that I don't want Meta accessing, and you do too. For me, it's nieces, nephews, medical bills, grievous leg wounds, funny vanity license plates, and probably thousands of embarrassing selfies I send my sister.
None of these are intended to be shared with a company that repeatedly steps on any available rake when it comes to privacy issues.
And just under a year ago, we learned that Meta Ray-Ban Display, Meta's take on smart glasses, has been used by many wearers to violate the privacy and safety of anyone around them. There's even an entire smart glasses subculture dedicated to disabling the lights that make it obvious when the wearers are filming.
I'm not entirely sure what Meta's end goal is, other than to figure out how to turn every single person into a perpetual data generation machine for its company.
And the punchline is, this whole Muse debacle is yet another trumpet blaring as we barrel full speed ahead into the oncoming AI privacy apocalypse. At what point is it too late to solve these problems?
If the agents are already blatantly ignoring the guardrails, are we already past the point of no return?
Your data is their business
I think what sticks in my craw about this the most is the fact that Meta has ensured that you do not get the option to opt out. No matter who you are, Meta has found a way to harvest your data.
Prompts to download Muse in WhatsApp, Instagram, and an ad for Muse in the iOS App Store
You could never create a single account on a single Meta service, of which there are many, and you still aren't safe. If you're talking to someone who has installed Muse, Meta could, presumably, access, read, analyze, and upload that information to its services.
And this is the point where I need to step back and point out that this isn't just a Meta issue. I'm naming-and-shaming Meta here because we've got countless examples of Meta's tape-and-bubblegum safety guards failing.
But every company that is investing in AI wants carte blanche access to your information. It doesn't matter if it's Meta, OpenAI, Anthropic, or even Apple.
Big tech companies want one thing, and it's disgusting.
I mean, hell, Apple Intelligence can already access things like Messages and Mail. The defining difference here is trust. Apple Intelligence, at present, attempts to process much of the request on-device.
That's a big reason why Apple Intelligence is relegated to more-powerful Apple chips.
In the event that something does need to be processed off-device, Apple says that only the relevant data is sent to Private Cloud Compute.
We're entering an era where privacy does not exist. At one point, the stock advice to people who were concerned about this very thing was "Well, if you don't like it, don't use it.
But that doesn't protect you anymore. Muse can get access to your information via people who do use it. Apple has made Apple Intelligence non-optional on compatible devices.
This is it, dear reader. You don't get to opt out anymore.