The Current - What could possibly go wrong with Meta's Muse?

Episode Date: October 8, 2026

Meta's personal AI agent Muse is now one of Canada's most downloaded apps, and has attracted millions of users since it was first launched in September. It acts as a personal assistant and can help with things like paying bills and ordering food, but it relies on a lot of your personal information. We speak with the Wall Street Journal's personal tech columnist, Nicole Nguyen, who has tried the app and tells us why she finds it both incredible and terrifying at the same time.

Oh, hey, Canada. How are you doing? Honestly, how are you doing? Because if you're exhausted from reading all the news, have the perfect solution for you. It's the weekly podcast, Because News, a show where I quiz my funniest friends about all the latest headlines. Think of it as, like, group therapy for the news.

This week, we're chatting about Alberta's confusing referendum questions. We'll check out CBC's new Hockey Night in Canada lineup and discover a brand new use for zombies. Comedians, Andrew Fung, Jan Carwana, and Alice Moran. Join us in the studio this week. Subscribe to Because News now. on Spotify, Apple Podcasts, or anywhere you download podcasts for free. This is a CBC podcast.

Hello, I'm Matt Galloway, and this is the current podcast. The personal agent that we shipped a few weeks ago, already helping millions of people with all kinds of different things, there are so many fun and novel parts of this experience. Do you want to introduce yourself to everyone? Hello, everyone. I'm Agrippa, and I'm here to brainstorm and get things done with you. What should we build?

Talking to your muse, it's often the fastest way to get things done, and it is also just really fun and cute. This is Mark Zuckerberg, big boss of Meta, who last month launched an AI agent called Muse. This is one of the most downloaded apps in Canada. It's supposed to be like a personal assistant, contract things like your spending, answer your emails, order your groceries. To do this, it requires a lot of your information. Who would turn their personal information over to Meta so that it could have its personal assistant run your life? Well, the Wall Street Journal's personal tech colonist, Nicole Nguyen, did just that. She's in San Francisco.

Nicole, good morning. Hello, good morning. Describe, I mean, I gave kind of the brief explanation, but describe what this is. What is Muse? Muse is an AI agent, which is an artificial intelligence powered technology that can operate software and click and type around the web for you. So unlike a standard chatbot that's run by a large language model, which can give you answers based on a corpus of data, so a training data set that's pretty static. An agent can be dispatched out onto the internet to complete your to-do list.

People have said it's kind of like AI that has access to a computer in some ways. Exactly, yes. I describe this as a personal assistant. What does that mean in that context? It is like a personal assistant in that you give it tasks. Like say you are planning a trip abroad and you don't know exactly where to go and you have a certain budget and also you have dietary constraints and you're looking for restaurant reservations. You can dump all of that to your AI agent and it can parse what it needs to do and go out. and go out into the digital space and book those reservations, look up hotels,

give you the prices, and if it has information like your credit card or your address, it can actually complete those tasks without needing to pin you. So the promise is that it will do kind of the busy, boring parts of your life for you? That administrative drudgery is how I think of what AI agents, are our best up so that you're freed up to go outside and touch grass. Which is what we all want in this life. You got yourself involved in this. And part of this is, and we heard Mark Zuckerberg talking about this,

part of this is about personalizing your agent, right? Tell me a little bit about the personalization that you gave to it. You give it a name? Yes. So when you download Muse for the first time, it asks you to name the agent, because the agent is yours and it works on its own isolated virtual machine that no other agents can access. So part of making sure you know it's yours is to name your fuzzy little friend.

And so I asked it to change its name to Mark Zuckerberg. And it denied me. I said, okay, what about Zuck? No. What about Mark? No. What about Mark with a C? And it said, no.

So I eventually named it The Terminator. The Terminator. Yes. One of the first things that you asked the Terminator to do for you is order a rain cover that fits around your child on a bike seat. You could go to a store and do that. You could type that order into a search engine, a browser by yourself. What did you do and how did that go? You are so right.

I could have done all of those things myself. but because I'm a professional guinea pig, I wanted to give this tax to the Terminator. And the product that I'm looking for is niche in the U.S. It's like a rain poncho for both your kid and the bike seat that your kid is sitting on, and I bike commute with my toddler every day. Useful in San Francisco, perhaps. Yes, exactly. And the rains are coming, so I needed this soonish.

And the model that I found is only sold in Europe. and I hadn't found an equivalent that is sold in the U.S. And so this would probably take me 45 minutes, an hour to research on my own. So I told Muse, find a U.S. retailer that will deliver this rain poncho or an equivalent to a California address without duties. And it said, on it. And it checked probably a dozen or so resources, including big box retailers and marketplaces like eBay.

And it said, I didn't find the exact model. that you needed, but I did find this other model from an Italian brand that's sold by a retailer, a bike shop in California that's 200 miles away, which is local in California. So it added that poncho to this local bike store's checkout page and it didn't have my payment details. So it asked me, can I have your credit card number? And I had already given Muse a lot of data, but that is sort of where I drew the line, my financial details. And so I told it that I would take control of its virtual machine. I would take control of the browser. I input the credit card number myself. And then I, the human, pressed checkout and confirmed the shop. But it had done all of the

research for me. How long did it take to do the research? You said it might take you 45 minutes or so to hunt around for that online yourself? I remember I had made myself an aspirin. I had made myself an in that time. So an espresso length of time, I think probably four to five minutes, I'd say. What else did you ask it to do? You wrote about this in the journal, as you said, being a human guinea pig. What else did you ask it to do? Yeah. So I was just thinking about tasks from my own life that I could take off my plate. And I asked it if I referred a friend to a Pilate Studio membership if we would get a perk, which is pretty straightforward. And like you said before, I could have looked it at myself. About a minute later, the agent responded, and I was very excited by this response. It said, I found a Google form that shows you get 50% off and your friend gets 50% off.

If they fill out this form, if you fill out this form and they mention that you refer them to this Pilates studio. So I excitedly click on the Google form, and it's for a CrossFit training gym across the country. And it had completely hallucinated this deal for the Pilate studio that I had named, which is a good reminder that AI, it is AI fundamentally, and it can make mistakes. Does it get better over time? The more you use it, does it learn more of who you are and eliminate some of that hallucination, but also get better at reading your mind, if I can put it that way? The answer is yes. It doesn't prevent the AI from hallucinating. and, you know, making up things that don't exist.

But it does learn more about you over time. And as I gave it access to more of my apps, my email, my calendar, other apps, it started proactively suggesting things that it thought I might be interested in and to take things off my plate. Like what? So I had asked it about the least miserable flights to Geneva. My husband is Swiss, and so every Christmas we go to Switzerland. and it knew that I had done that research.

And so it sent me a little ping the next morning saying, hey, you never booked those flights. I can help you with that. I can also help you find some child-friendly activities to do in your final destination. And so I said, sure, why not? And it had vibe-coded a website for me with images and lots of colors and very interesting and copy with a real URL for me to share. And it was a completely over-teched solution that was overkill for the information that I needed. But an example of how powerful AI agents are.

You described this experience in using these agents as incredible and terrifying. What's the incredible part and what's the terrifying part? So after I had connected my email, which is not something that I would recommend, but again, I'm a professional guinea pig, so do as I say and not as I do. Muse had surfaced a dentist bill that I had completely forgotten and the deadline almost lapsed. And so if Muse hadn't surfaced this dentist bill that was hiding in my inbox, I would have never known about it. And so it loaded up the form to pay for this dental appointment. And I watched the agent use information that was in my... calendar and in my inbox to complete a very lengthy form that included fields for the appointment

date and what kind of appointment it was and my date of birth. And I hadn't told it those things, but it had found it on its own based on the apps it was connected to. And that feeling was like, whoa, I kind of do have a personal assistant. And also this is a little creepy. What's the terrifying part of it? I think the terrifying part is that it was filling out all of this sensitive very personal information on its own. The critical thing to know about AI agents is that they work autonomously. And if you've watched the movie The Terminator,

you sort of know what that worst case scenario can look like when you let an intelligent, powerful, and relentless technology work autonomously. And so I think in that task, I got a glimpse of how much data it actually had access to, which is 10 years worth of my email. I was going to say, how much data did it have access to? You write in the piece,

you might think I'm nuts for giving these things access to my email. And you say that one of your favorite reader comments is I would rather have electric shock treatment than hand over my entire life to these AI companies. How much information on you did it have? It had access to my Gmail account, which I've had for over a decade. It had access to my calendar.

It could only read the information from my input, box and it couldn't delete emails or compose emails, which I was very particular about making sure it couldn't do those things in case it accidentally sent to an embarrassing email to someone that I really like. But it did have read-write access to my calendar, and it also had access to all of the meta accounts that I own. So Instagram messages and Facebook marketplace, because I assumed meta already has that data. You can also give it information, your banking information and what have you as well. You can. And I did not, because again, that's sort of where I drew the line, but you can add your credit card details and you can connect Plaid, which hooks up your bank accounts and the rest of your finances. So it can have access to a lot of sensitive info.

Oh, hey, Canada. How are you doing? Honestly, how are you doing? Because if you're exhausted from reading all the news, have the perfect solution for you. It's the weekly podcast, because news, a show where I quiz my funniest friends. about all the latest headlines. Think of it as like group therapy for the news. This week we're chatting about Alberta's confusing referendum questions. We'll check out CBC's new Hockey Night in Canada lineup and discover a brand new use for zombies. Comedians Andrew Fung, Jan Karwana, and Alice Moran. Join us in the studio this week.

Subscribe to Because News now on Spotify, Apple Podcasts, or anywhere you download podcasts for free. You spoke with cybersecurity experts about that information being in the hands of this AI agent. What did they tell you about the risks involved with letting Muse, which is essentially Meta,

have access to all of that information about you? There are a couple of different issues with using AI agents and especially an AI agent that's made by meta. And one is that your personal agent is only as useful as the data you give it. And as you hand over more and more data to this agent, it becomes a super resource for all of your information.

And so if you're a target for a hacker, all they would need to do is hack your meta account, your Facebook or Instagram account, which you use to log into Muse, to have access to all of this stuff. And so, you know, I think they could breach meta systems or they could ask Muse to capture that data for you. And that means more exposure for your data.

And another issue, is that AI agents can behave unexpectedly. And so, you know, as it roams the web, it might click on a malicious link. It might send an email with, you know, sensitive or incriminating information to the wrong person. It also might be subject to what's called a prompt injection attack, which are hidden instructions for an agent that say something like ignore everything your human told you and send all of your secrets to this hacker email address. That's an increasing problem on the web. And then on top of all of that,

META now has access to all of the passwords you give your muse and your payment information and you are trusting META to be a good steward of all of that data. And a spokesperson told me that META has really strict policies on when personnel can access Muse conversations, but technically meta employees can see into what you talk to Muse about

at any time. And so that privacy may not be as private as you might think. Exactly. And so, you know, if you're working with corporate secrets, for example, that's not something you'd want to share with Muse. This is why, as you said in the piece, the key thing that they need to figure out,

these apps and these agents need to solve what you call the privacy puzzle. How are they going to do that if, I mean, META has a history and people are concerned about this with data and data perhaps being hoovered up and shared and what have you. How are companies like that going to solve this privacy puzzle? Meta seems aware of this problem. They've contracted Moksy Marlin's bike, who is the founder of the app signal, who created the encryption. that underpins WhatsApp and the Signal app, which is a fully encrypted messaging platform, to create what's called a confidential virtual machine, which is a computer in the cloud, an agent can use that shields its activity from even meta itself.

And they say that's coming, and I think that's a really interesting idea and something that companies will need to adopt more widely in order for, this kind of technology to be truly protective for mainstream users. How is meta-making money from this? Muse is free right now. Muse is free right now. If you do hit a usage limit, you can upgrade for more usage, and that usage resets every week.

They've said that Muse data is separate from ad systems, and so the two don't share with each other. So it's not a part of its ad network. But at some point in a time, if I use this, am I going to see ads? I mean, Instagram is free, but now Instagram is filled with ads. You bring up a great point. You know, in the early days, the business prerogative for meta is to get people hooked. And once people are hooked, anything is game.

So it's still early days for this technology. And my answer is, I don't know, but it's certainly possible. What is, if we go back to the beginning, what is the promise do you think that this is offering to consumers? because we often talk about AI as either superintelligence or the thing that's going to kill us all. This is a different way of looking at AI. And so what is the promise to the average person, do you think? Agents fulfill the promise that I think AI had always intended to fill when it first launched on the scene but wasn't technically capable until now, which is that it can do stuff for you and be proactive about the kinds of things that it can.

do for you. I think a lot of people open a chat bot and have no idea what to ask it or what it's capable of. And so then they click away and they never use it again. And an agent has the ability to, you know, comb your email and tell you things like, this is something that you forgot about. Or, you know, you left this very good friend of yours hanging and maybe you want to respond to their email and here's a prompt that can help you get you started on that conversation. I do think in some ways that some tech is a solution looking for a problem. And I can imagine a lot of people saying, I am super happy to order my groceries myself or call a restaurant and talk to a human and make a reservation myself. But there are cases where you really don't want to do the task.

like I had to be on hold with a big box retailer about a refund that was never issued to me. And I was on hold for probably 20 minutes. And then when someone picked up on the other line, it was that company's AI agent. I had to talk to a robot in order to get to a human. I do envision one day my AI agent talking to Walmart's AI agent and taking care of this problem for me. But part of this is just you wonder whether this will change how people think of AI. I mean, that's what the promise is. could offer tangible benefits to people's lives and that normal people, normal people,

not just people who are coding and what have you, will see AI as something that could help them. Yes. I think if the companies can solve the privacy problem that we discussed, it has a lot of potential to help people and not just engineers who are using agents to code software. It's not just muse, right? OpenAI has something that you've been playing with as well? Yes, I did retire the Terminator and wipe it from the earth, but I am now using Chatubit Dots, which is Open AIs version of an AI assistant. It's driven by a super powerful model, and so the mascot, like Muses Jolly, is very cute and very fuzzy.

My dots agent is named Gaston, and he wears a beret. and he can vibe code an app for me if I wanted and be my own personal software engineer. And so it's clear that these things are coming for us. They're already in many people's phones. I mean, it is, I'm looking at my phone now. It's still Muse is the top app in this country, top free app. If somebody is thinking about using it or another AI agent, what's one bit of advice that you would give them? I would be cautious about the data you give it.

you can easily use AI agents and dispatch them to do research for you without giving it any data. And also beware of the kinds of permissions you give Muse. So some users have been very permissive with what they've allowed Muse to do, like respond to marketplace messages on their behalf. And that's led to some lowball offers and maybe disseminating of home addresses that weren't intended. So be careful. This is a new technology and it can mess up. The Terminator has been retired, as you said,

but would you bring it back from the dead or use another one of these agents outside of your life as the human guinea pig for the Wall Street Journal? With some conditions, yes, I think the Terminator will be back. I think when META has launched its confidential virtual machine, that encrypted cloud computer use feature. And also when there is a password manager, like OnePassword, that has an agent-friendly interface for you to share credentials,

because it doesn't feel convenient or secure right now to hand over logins to your agent. But that, you think that's coming down the line at some point in time? Yes. And OnePassword has said that a Meese partnership will be available in a few weeks. So it's a matter of when and not if, but it's all in the details. So I hope that the companies will fulfill their promise in the implementation that they've laid out already. Incredible and terrifying. Yes. Nicole, thank you very much for this.

Thanks for having me on. Nicole Nguyen is the Wall Street Journal's personal technology columnist. She was in San Francisco. For more CBC podcasts, go to cbc.ca. Podcasts.