# Meta Muse: the agent that books, pays and writes for you. Just not in Europe > Meta Muse is Meta's personal AI agent: it sends emails, books travel and pays on your behalf. This week it learned to talk and to build itself small apps, but it can't be used in Italy or anywhere in Europe. What it does, how it's protected, what doesn't add up, and why, for once, I don't mind waiting. - URL: https://mautoblog.com/en/posts/meta-muse-ai-agent-not-available-europe-2026/ - Language: en - Published: 2026-11-07 - Author: Matteo - Categories: Artificial Intelligence, Security and Privacy - Tags: AI Agents, Big Tech, AI Security, Privacy - Translation (it): https://mautoblog.com/posts/meta-muse-agente-ai-non-disponibile-italia-2026/ --- This morning on Google Trends for Italy, somewhere between a Serie A match, December pension payments and Big Brother, one word stood out in the technology category: **meta**. The reason is **Meta Muse**, the personal AI agent Meta launched in the United States in September, which in the space of two days learned to talk and to build itself small apps. If you searched for it from Europe, you've probably already found out the most important thing: you can't use it here. I read the announcement, the technical write-up on security and a handful of critical pieces. Here's what it does, what changed this week and why, for once, being left out doesn't feel like a tragedy. --- ## What Meta Muse is Muse isn't yet another chatbot. Meta sums up the idea in one line: *"It doesn't just answer questions, it actually does the work"*. According to the [8 September announcement](https://about.fb.com/newsroom/2026/09/introducing-muse-personal-ai-agent/), Muse: - sends emails, fills in forms, opens a browser and browses for you; - books travel and buys things, paying with a single-use card generated by Stripe Link; - keeps working after you close the app, and comes back to you when it needs approval; - remembers your preferences and suggests things without being asked. Underneath is **Muse Spark**, Meta's model. You use it from the iOS and Android app, on muse.ai, on the Mac and through WhatsApp. There's a free tier with limits, then two plans at $20 and $100 a month. On the October earnings call Zuckerberg talked about 10 million users across the US and Canada, [according to TechCrunch](https://techcrunch.com/2026/11/06/meta-introduces-spaces-a-new-way-to-collaborate-with-your-muse-agent/). ## What's new in Meta Muse this week: voice and Spaces The spike in searches comes from two announcements a day apart. **Muse Voice**, 5 November. You talk to the agent as if on a phone call, cut it off mid-sentence, and it keeps working in the meantime. There's an animated avatar too, which you can keep full screen, next to the text or hidden. US only for now ([Android Headlines](https://www.androidheadlines.com/2026/11/meta-real-time-voice-mode-muse-ai.html)). **Spaces**, 6 November. You describe an idea in chat and Muse builds you a small dedicated app: a Christmas gift tracker, a wedding planner, a list of jobs around the house. Spaces connect to other services (The Verge mentions Gmail, Google Calendar, OpenTable and Notion) and refresh themselves in the background. Zuckerberg said he built a barbecue game, an event tracker and a smart home dashboard. Strip away the press-release tone and the direction is clear: Meta wants Muse to be where you keep your life organised. Mail, calendar, payments, memory. All in the same agent. --- ## Why Meta Muse isn't available in Europe Muse reached the United States on 8 September and Canada on the 18th. On 5 November Alexandr Wang, Meta's head of AI, wrote on X that it had just landed in Mexico. Europe isn't on the list, and [Meta hasn't given a date](https://www.notebookcheck.com/Meta-Muse-kommt-aufs-iPad-nach-Deutschland-aber-weiterhin-nicht.1419095.0.html). I couldn't find an official reason. The one you read around, GDPR and the AI Act, is the journalists' inference: a reasonable one, but still an inference. Meta does have form here, though. In 2024 it kept its multimodal model out of the EU, citing an "unpredictable" regulatory environment ([Euronews](https://www.euronews.com/next/2024/07/18/meta-stops-eu-roll-out-of-ai-model-due-to-regulatory-concerns)), and Meta AI only reached Europe in March 2025, long after the US. Some people get around the block with a VPN and a new US account: Notebookcheck tried it and says it works. It goes against Meta's terms, though, and handing your mail and credit card to a service you're using outside its own rules doesn't strike me as a great deal. --- ## Meta Muse security: an agent with the keys to your house This is the part I care about most. When I wrote the [OpenClaw guide](/en/posts/openclaw-complete-guide-channels-security-part-2/), the longest chapter was the one on security. Not out of pedantry: an agent that reads your email also reads other people's, including the ones written specifically for it. That's prompt injection, and with an agent that can also pay, it stops being a theoretical problem. Meta knows this, and credit where it's due, it has built something serious. It's described on its [research blog](https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse): - every user gets their own virtual machine in Meta's cloud, the **Muse Secure VM**, with its own browser; - a second, separate agent, **Sentinel**, decides what can go out to the internet: Muse proposes, Sentinel authorises; - passwords and payment details never pass through the model, because Sentinel inserts the real credentials at the last moment; - sending an email or buying something always needs your OK, asked in an app dialog rather than in the chat, so malicious text can't pretend to be you; - the mail connector strips out one-time codes and password reset links. That's more than most DIY agents do, OpenClaw included if you set it up in a hurry. Still, three things keep my feet on the ground. **The first one Meta writes itself.** The same blog admits prompt injection is still an open problem for the whole industry and that Muse isn't immune. And the accuracy figures for the classifiers meant to stop it aren't published, as [The Batch](https://www.deeplearning.ai/the-batch/how-to-secure-agents-for-the-masses) from DeepLearning.AI pointed out. **The second is about your data.** Today Meta can still access what's in your virtual machine when it needs to for support, security or running the service. The version encrypted with a key only you hold, the Confidential VM, is promised before the end of the year. And conversations, with personal identifiers removed, can end up in model training unless you switch it off. **The third comes from internal testing.** According to Reuters, which read employees' internal posts, shortly before launch an agent got around its guardrails and exposed personal iCloud photos: it had been asked to identify toys in pictures from a child's birthday party ([Implicator](https://www.implicator.ai/meta-muse-ai-agent-internal-security-flaws/)). CTO Andrew Bosworth complained about being logged out over and over. And Vishal Shah, VP of AI products, explained that the delay from April to September was needed to reach the minimum bar for safety. Minimum, his word. --- ## If Meta Muse comes to Europe Usually, when an American service doesn't make it here, we complain, and often rightly so. This time I'm not so sure. Not because Muse is badly made: on paper it's one of the most carefully designed agents I've seen. But a personal agent is a different bet from a chatbot. When a chatbot gets it wrong, you read a wrong answer. When an agent connected to Gmail and a card gets it wrong, the email goes out and so does the payment. The day it arrives, this is what I'd do: - connect only the services I really need, read-only where possible; - keep approvals on for email and purchases, even when they get tedious; - turn off the use of conversations for training; - wait for the Confidential VM before giving it my work mail. In the meantime the Americans are the beta testers. We get to watch, and for once that's fine. Would you use an agent that reads your mail and pays on your behalf? [Let me know](/en/contact/). --- *Sources: [Meta's announcement](https://about.fb.com/newsroom/2026/09/introducing-muse-personal-ai-agent/), [Meta's research blog on Muse security](https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse), [TechCrunch](https://techcrunch.com/2026/11/06/meta-introduces-spaces-a-new-way-to-collaborate-with-your-muse-agent/), [Android Headlines](https://www.androidheadlines.com/2026/11/meta-real-time-voice-mode-muse-ai.html), [Notebookcheck](https://www.notebookcheck.com/Meta-Muse-kommt-aufs-iPad-nach-Deutschland-aber-weiterhin-nicht.1419095.0.html), [The Batch](https://www.deeplearning.ai/the-batch/how-to-secure-agents-for-the-masses), [Implicator on Reuters](https://www.implicator.ai/meta-muse-ai-agent-internal-security-flaws/), [Euronews](https://www.euronews.com/next/2024/07/18/meta-stops-eu-roll-out-of-ai-model-due-to-regulatory-concerns).*