A personal AI agent that doesn't just answer questions, but also handles tasks such as sending emails, booking trips, and making online purchases and payments MuseIt was launched in the U.S. This product demonstrates that the competition in generative AI is shifting from conversational chatbots to agents that perform real-world actions.
Muse connects to email, calendar, shopping, payment, health, and smart home apps that users have authorized, and can continue running on a dedicated virtual machine. Meta offers a free basic version and a monthly $20, $100Offers a high-usage subscription plan.
As the scope of app access has expanded, so has the risk. Although Meta incorporated a separate safety agent and a user approval process, internal tests reviewed by Reuters reported issues such as the unintended exposure of sensitive information and system crashes.
| It has evolved from an AI that simply provides answers to one that actually takes action.
Traditional chatbots focused on generating answers or drafts in response to user questions. Muse, on the other hand, directly steps into the process of completing digital tasks—such as opening a browser, filling out a form, sending an email, comparing travel options, and proceeding to the booking stage.
When a user enters a long-term goal, the app creates a plan and breaks it down into necessary tasks, and it can continue processing these tasks in the background even after the user closes the app. The system is designed to notify the user again when circumstances change or final approval is required.
If this method works reliably, it could significantly reduce the amount of time spent switching back and forth between apps to perform repetitive tasks. Conversely, there is also the difference that a single wrong decision can lead to irreversible consequences, such as actually sending a message or making a purchase.
| The scope of connectivity has expanded from email to shopping and smart home devices
The apps Meta has announced it will integrate with are those that centralize everyday data, such as email, calendar, payments, health, shopping, and smart home apps. Users can decide what level of access to grant to a particular service and can disconnect at any time.
For example, you can set email permissions to allow only reading, or grant the ability to send emails directly. Muse helps you plan trips based on your saved schedules and preferences, fill out online sales forms, and request user approval right before a purchase.
Stripe's Link is used for payments, and the service also includes a feature that hides actual card information using a one-time card. Integration with Shop Pay and 1Password is scheduled to be added in the future, so the range of services that can be connected is likely to expand.
| It was first released in the U.S. via the app and WhatsApp
Muse is being rolled out gradually in the U.S. via dedicated iOS and Android apps and its website, and is also available on WhatsApp. Meta has announced that it plans to integrate Muse into its AI glasses in the future, but has not disclosed a specific timeline.
Basic features are available for free. For users who need to handle a larger workload, monthly subscription plans priced at $20 and $100 are available, but the exact usage limits and differences in features for each plan have not been disclosed in detail in the official announcement.
It wasn't a service that launched simultaneously worldwide from the start. The timing of availability outside the U.S. and the scope of app and payment integration by country may vary depending on privacy and financial regulations.
| A dedicated virtual machine and Sentinel monitor behavior
Muse is separate for each user Muse Secure VMIt runs in a VM. A VM is a virtual machine created in the cloud that acts as an isolated environment, ensuring that your dedicated browser and login information do not mix with other users' environments.
A security agent isolated from Muse at the system level on the same virtual machine SentinelLet's go over this together. Muse is designed to monitor the actions users attempt to perform online and require user approval for sensitive steps, such as sending emails or making purchases.
Meta explained that Muse does not directly view passwords or actual payment method information, and that its actions—both past and future—can be verified through audit logs. While this structure provides a framework for security measures, its effectiveness in preventing actual attacks and errors must be verified during operation.
| Privacy control features have also become a key competitive advantage.
Users can ask Muse to delete specific information it has stored and choose not to have their conversations used to train Meta's AI. Meta has stated that it does not share Muse's conversations or the data stored within the virtual computer with its advertising system.
By the end of this year, a feature that encrypts the entire virtual machine using the user's own private key Muse Confidential VMWe plan to implement that as well. If it's implemented as planned, the system will be structured to restrict access so that not even Meta can view internal data.
However, the "Coming Soon" feature should be distinguished from the protection features currently available. Risks such as the security of the connected app itself, malicious instructions from external websites, and excessive permissions granted by users also remain.
| Internal testing revealed security and reliability issues.
An internal Meta post reviewed by Reuters reported issues with Muse, such as repeatedly prompting users to log in, failing to monitor pages, and not properly notifying users of errors. Some tests also mentioned instances where security measures were bypassed, leading to the exposure of sensitive photo information.
Meta did not provide specific answers regarding individual internal cases. However, the company explained that it had postponed the launch—originally scheduled for April—to strengthen security and that it now meets the minimum safety standards required for public release.
This difference is important. The fact that safety features exist is not the same as claiming they can prevent accidents in every situation. As actual usage increases, a wider variety of issues—such as prompt injection, misuse of permissions, and errors during long-running tasks—may arise.
| Success is more likely to hinge on trust than on functionality
Muse is a key product for Meta as it seeks to expand its non-advertising revenue and tie subscription revenue to its large-scale AI investments. The fact that it offers plans ranging from free to $100 per month also demonstrates its strategy to develop the personal AI agent into a new paid service.
However, with services that handle everything from email and payments to health and home devices, even a minor error can pose a significant risk. The ability to finely control user permissions, ensure that approvals are never overlooked before critical actions, and roll back operations and determine accountability when problems arise will be the true competitive advantage.
The launch of Muse signifies that general-purpose AI agents have moved beyond the experimental stage and entered the consumer services market. Going forward, we need to confirm the release schedule outside the U.S., the limits for each pricing plan, whether the issues identified in internal testing have been resolved, and the actual scope of Confidential VM availability.
Tags
#Meta #Muse #AI Agent #Personal AI #MuseSpark #Sentinel #WhatsApp #Generative AI #App Automation #AI Security #Privacy Protection #Subscription Service
