I tried Meta’s new Muse AI agent — it’s incredibly useful, but handing it my digital life felt deeply uncomfortable

Meta’s new Muse AI platform would like the keys to your digital life to help run it for you.

<![CDATA[ <article> <p>Meta’s new Muse AI platform would like the keys to your digital life to help run it for you. Muse is a personal AI agent that <a href="https://www.techradar.com/ai-platforms-assistants/meta-thought-ai-could-do-the-job-instead-of-humans-what-happened-next-should-surprise-absolutely-nobody">Meta</a> built to complete tasks on your behalf, including sending emails, filling out forms, shopping, and pretty much any other activity with an online component. </p><p>The AI is available in the U.S. as an app, or through <a href="https://www.techradar.com/uk/tag/whatsapp">WhatsApp</a> or the <a href="https://muse.ai/" target="_blank">Muse.ai portal</a>. </p><p>It's a familiar idea in some respects. Muse matches OpenAI and other AI developers in being able to open a browser inside its own cloud-based virtual computer and move between connected services. It returns when it has finished, encounters a problem, or needs permission to do something sensitive. </p><p>Meta describes the experience as having an agent that understands your goals and advances them independently. That can come off as sinister, even before it starts asking for permission to essentially be you online by having you sign in to all your online accounts. </p><p>Getting started is straightforward, although Muse is currently limited to U.S. adults. You can install the Muse app from your phone's app store, open Meta's Muse website or access it through WhatsApp. There is a free usage tier, while heavier users can pay $20 per month for Power or $100 for Maximum. </p><h2 id="muse-shops-for-you">Muse shops for you</h2><figure class="van-image-figure inline-layout" data-bordeaux-image-check ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1920px;"><p class="vanilla-image-block" style="padding-top:56.25%;"><img id="q3cL5ag9F4ejbNtAXHwnhY" name="Meta Muse" alt="Meta Muse" src="https://cdn.mos.cms.futurecdn.net/q3cL5ag9F4ejbNtAXHwnhY.png" mos="" align="middle" fullscreen="" width="1920" height="1080" attribution="" endorsement="" class="inline"></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Meta Muse)</span></figcaption></figure><p>Once you open the app, the first thing the AI does is suggest you pick a nickname for it so you know when it is messaging you. I went with one of the suggested choices, and "Scout" proceeded to ask me what I'd like to do with it, offering a handful of popular options.</p><p>I asked Scout to do some shopping for me first, as the AI made it seem like a much smoother process than with other AI tools. I asked it to find me a pair of shorts. It asked my size and went off on its own for a couple of minutes, coming back with a couple of options and pointing to a sale on one particular set. I said okay, and it asked for my Amazon login to make the purchase on my behalf. For an everyday purchase, it was surprisingly pleasant to delegate.</p><p>That little interruption was reassuring and unsettling in almost equal measure. Meta says credentials entered this way are routed directly into secure credential storage and are not visible to the main Muse agent. Once stored, they can be inserted into the browser when needed without exposing the actual credentials to Muse's browser subagent. Muse does the legwork, while I remained the person responsible for deciding whether money actually left my account.</p><p>This is probably the easiest Muse task I would recommend trying yourself. Give it a product, let it browse, and take over whenever a login or other sensitive interaction appears. It demonstrates the agent's capabilities without immediately granting it access to years of personal correspondence.</p><p>Naturally, that was exactly what I did next.</p><h2 id="getting-muse-to-handle-my-email">Getting Muse to handle my email</h2><figure class="van-image-figure inline-layout" data-bordeaux-image-check ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:1000px;"><p class="vanilla-image-block" style="padding-top:56.30%;"><img id="Z3qkzoGTB9FTZ8VMKj9c2K" name="shutterstock_408362719.jpg" alt="Email warning" src="https://cdn.mos.cms.futurecdn.net/Z3qkzoGTB9FTZ8VMKj9c2K.jpg" mos="" align="middle" fullscreen="" width="1000" height="563" attribution="" endorsement="" class="inline"></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Shutterstock)</span></figcaption></figure><p>I took the plunge and told Scout I wanted help with my email. I had to give it permission to go through my inbox and to write on my behalf, and it asked to use it to learn about me. With some trepidation, I okayed that, and it put together the kind of distorted image of my life you'd expect from someone basing their impression of you solely on your emails. </p><p>Scout also brought up an unopened email from the power company about needing to check my solar panels. It offered to write a reply if I could pick from the dates suggested for when they would visit. I did, and a minute later there was a draft pending my approval. It looked fine, and suddenly another little online errand was done. </p><p>It felt different from pasting an email into an AI chatbot. Muse's AI could retrieve the context itself rather than making me act as a courier between my inbox and the AI. Of course, it also felt a lot more personal. Meta has acknowledged how sensitive this access is and built specific defenses to prevent issues. </p><p>Muse's email connector filters out one-time tokens, password-reset links, and login links, while its Sentinel security layer controls what actions the agent is permitted to take outside its virtual machine. When an action requires my approval, the request comes through Muse's interface rather than simply relying on the agent to interpret a conversational yes. </p><p>I would keep that friction in place for email, particularly while getting used to Muse. Saving three minutes responding to a utility company is nice, but it is not worth discovering that your AI assistant has developed an unexpectedly adventurous correspondence style.</p><h2 id="muse-is-convenience-mixed-with-discomfort">Muse is convenience mixed with discomfort</h2><figure class="van-image-figure inline-layout" data-bordeaux-image-check ><div class='image-full-width-wrapper'><div class='image-widthsetter' style="max-width:970px;"><p class="vanilla-image-block" style="padding-top:56.19%;"><img id="2BaNK5XKNiUsUgc3MA8WBC" name="AI security" alt="AI security" src="https://cdn.mos.cms.futurecdn.net/2BaNK5XKNiUsUgc3MA8WBC.jpg" mos="" align="middle" fullscreen="" width="970" height="545" attribution="" endorsement="" class="inline"></p></div></div><figcaption itemprop="caption description" class=" inline-layout"><span class="credit" itemprop="copyrightHolder">(Image credit: Pixabay)</span></figcaption></figure><p>The more I engaged with Muse, the more the tandem senses of convenience and discomfort grew. Muse feels more frictionless than similar services with virtual browsers. And the combination of data Meta already has and the data it required permission to access made it feel more personal, too. Every improvement in its ability to help me came with another small moment where I had to decide how much access I was willing to grant.</p><p>Meta says conversations and VM data are not shared with its advertising systems, although browsing activity performed on your behalf can still influence advertising indirectly because websites may treat Muse's visits as your activity. Meta says sanitized agent interactions and tool activity can be used to train future models, but users can opt out with a switch in Muse's settings. Meta is also developing a Confidential VM mode intended to cryptographically prevent even Meta from accessing information inside the VM.</p><p>I was more impressed with Muse than I expected. But that only made me warier about it. Digital chores that consume more time than any one of them deserves are all too common at the moment, and Muse proved remarkably good at stepping into that gap. The catch is that the closer it gets to behaving like a genuinely useful personal assistant, the more personal information it needs to do the job.</p><p>If you're thinking of trying Muse then I would recommend starting gradually, beginning with public web tasks and keeping consequential actions behind approvals at first. The frustrating part is that Muse makes a very persuasive argument for personal AI agents. Meta has built something that makes handing an AI more of my digital life appealing, but not when it's attached to the current crop of tech giants offering those services. </p> </article> ]]>