Updated September 10, 2026. Rollout has begun in the US; Italian availability has not yet been announced.
Meta Muse AI is the new personal agent presented by Meta on September 8th. Unlike a chatbot that responds and waits for another message, Muse can continue working in the background, use a browser, fill out forms, prepare emails, organize trips and ask for user approval before sensitive actions. The service debuts in the United States on iOS, Android and web; Meta did not indicate a date for Italy.
What is Meta Muse AI
Muse receives a goal, turns it into a plan and coordinates activities that can last over time. Meta mentions bookings, expense negotiation, selling a car, personal planning and purchasing. The underlying model is called Muse Spark and is described by the company as designed for real-world agentic work. Independent performance, however, is not yet available to a sufficient extent to compare it with other agents.
The new element is not the chat, but the ability to act. Muse can open pages, use connectors, generate tools, and coordinate sub-agents. This autonomy increases the usefulness but also the potential damage of an error. This is why Meta built a separate environment instead of connecting the model directly to the user’s credentials.
How Muse Secure VM works
Each user has a Muse Secure VM, a dedicated virtual computer in the cloud. The main process runs in an isolated container and does not directly see passwords or tokens. Credentials remain in a separate store and are used by components with limited privileges. The exit to the Internet and the actions on connected services pass through Sentinel, an independent agent that can block a request or ask for confirmation.
Meta says that the user can choose which applications to connect and distinguish, for example, between read access and the ability to send emails. A history of operations is available. A Confidential VM encrypted with a user-controlled key should arrive later, but this protection does not coincide with the configuration available at launch.
| Element | Launch situation |
|---|---|
| Availability | United States, iOS, Android and web |
| Italy | No official date |
| Model | Muse Spark |
| Environment | Dedicated cloud VM |
| Sensitive actions | They require approval |
| Price | Free basic use, advanced plans available |
Purchases, emails and WhatsApp
Muse can be used from your own application or via WhatsApp. For payments, Meta integrates Link from Stripe, which generates a single-use card and hides the real data; Shop Pay is expected later. 1Password support should allow the use of existing accounts without showing the password to the model. Confirmations are promised before sending messages or completing purchases.
These protections reduce the risk, but do not eliminate interpretation errors, malicious sites or prompt injections contained in the pages read by the agent. Meta recognizes that Muse will continue to make mistakes and has opened a bug bounty that can reach up to $300,000. It’s a useful signal, not an absolute guarantee.
Privacy and personal data
The VM contains conversations, memory, and data from connected services. Meta claims that this content is not shared with advertising systems and that the user can exclude interactions from training. Service conditions, necessary telemetry and external connector rules remain important. Before entrusting sensitive emails or documents, it is best to check the options actually available in your country.
The security model “assumes compromise”: the agentic core is treated as software that can be deceived. This is a more realistic approach than simply promising that the model will always follow instructions. Our insights intooperational risks of AI agentsexplains why minimal permissions and isolated environments are essential.
When Meta Muse AI arrives in Italy
As of September 10, Meta has not communicated the European calendar, supported languages or subscription prices. The presence of personal data, storage, payments and autonomous actions may require regulatory adaptations before launch in the European Union. Any site that already indicates a precise Italian date is therefore presenting a prediction, not official information.
Meta Muse AI opens a phase in which the assistant does not simply suggest, but operates. The crucial test will be to see how many tasks he completes correctly, how often he asks for confirmation and how he reacts to hostile content. Until independent testing, claimed features should be considered a product roadmap, not a definitive demonstration of reliability.
We will monitor Meta Muse AI over the coming weeks and update this page when independent data, final terms or availability changes emerge. For anyone evaluating it today, the essential rule is to separate accessible features, vendor claims and results that can be reproduced independently.
For further context, read this related CryptoRoad analysis.
