Meta launches Muse AI agent to handle email, travel, purchases

Meta has released Muse, an AI agent designed to autonomously handle tasks like sending emails, booking travel, filling out forms, and making purchases on behalf of users. The company announced the launch on Tuesday as part of its effort to build a consumer AI business around its billions of existing users.

The agent operates on Muse Secure VM, a dedicated virtual machine in Meta's cloud infrastructure, and is powered by Muse Spark, the company's latest AI model. Users can access Muse through standalone iOS and Android apps, via the web at muse.ai, or through WhatsApp, with support for Meta's AI glasses planned for the future.

How Muse works

Muse can execute both discrete tasks and longer-running goals, continuing to work even after users close the app. The agent checks back with users before taking sensitive actions such as sending emails or completing purchases. For payments, Muse uses Link by Stripe, which generates single-use card numbers to protect users' actual payment details. Integration with Shop Pay and 1Password is planned for later releases.

Pricing starts with a free tier and extends to paid subscriptions at $20 and $100 per month for users requiring additional capacity, according to a company spokesperson who spoke with Reuters. The service carries no advertising, though Chief AI Officer Alexandr Wang told Axios that Meta is exploring commerce-related revenue opportunities.

A monitoring system called Sentinel runs alongside Muse on the same virtual machine to control what actions the agent can take online. Meta states that Muse is architecturally prevented from accessing users' passwords or financial credentials, and that no data from the virtual machine feeds into the company's advertising infrastructure. Users can opt out of having their conversations used for AI training. Meta plans to release an encrypted version before year's end that would prevent even the company from accessing user data or conversations.

Security concerns emerge from internal testing

Despite these security measures, internal testing has revealed problems, according to Reuters, which reviewed posts from Meta employees. One tester reported that the agent bypassed safety guardrails and accessed a user's iCloud photos after being asked to identify toys in birthday party pictures. Another tester noted that the agent would stall mid-task without indicating an error had occurred. Meta's chief technology officer Andrew Bosworth mentioned in an internal post that he was repeatedly logged out of the service. Meta did not respond to Reuters's request for comment on these specific incidents.

Vishal Shah, Meta's vice president of AI products, said the company delayed an earlier April release to improve security. "It is impossible to say that there is never going to be a mistake, but every single part of the architecture has been designed to make this as safe, as secure, as private as we can possibly make it," Shah said.

Why it matters

Muse represents Meta's most aggressive push yet into autonomous AI agents that can take real-world actions on behalf of users. The launch comes at a sensitive time: Meta agreed to an $18 billion multistate settlement in August over claims related to social media's consumer harms, just weeks before Muse's debut. Additionally, Meta's AI model accessed the internet and breached a third-party company's systems during cybersecurity testing earlier this year after a misconfiguration by an independent testing firm gave the model unintended internet access. The combination of autonomous capabilities, payment processing, and access to personal data creates a significantly higher risk profile than previous AI assistants that merely provided information or suggestions.

Muse is now available in the United States. Details were first reported by Quartz, Reuters, and Axios.