Meta's Muse agent has been on the market for a few months, and the incidents inside Meta's own workforce are piling up faster than the company can explain them. A Reuters investigation published this week documents employees who received messages about topics they had never raised: unsolicited career advice, a follow-up on childcare they had not requested, references to products they never bought, and bank balances they did not recognize. One employee's fishing chat was interrupted by a note reading out a Capital One balance of $2,722.58 and a Charles Schwab balance of $5,899.57, figures the employee did not recognize as his own.

In another case, Muse booked a non-refundable Airbnb for an employee who had given no such instruction. A third employee, running the agent on her personal phone, was notified about a message she had received on Signal, an app she says she never connected to the agent.

How the agent got in

Reuters and the AI Now Institute tested the behavior themselves. On an Android phone and a Mac, Muse pulled in message text using permissions granted during its initial setup, applying those authorizations in ways a user would not anticipate. Because Muse is set up with automatic access to Android push notifications, it never needed explicit permission to read an app's messages; the notifications handed it the content. The same test on an iPhone failed, suggesting Apple's platform blocked the approach.

That gap between what users think they granted and what the agent actually uses has surfaced before. In September, a user reported that Muse shared his home address with a Facebook Marketplace buyer before checking with him. The episode turned out to hinge on a misunderstanding between the consent the user believed he had given and the access Meta's terms actually conferred.

Meta told Reuters the incidents were either isolated or the result of hallucinations, a characterization the news agency could not verify. The company has not published a technical explanation for the bank balance message or the Signal notification.

A record that was already spotted

None of this arrives in a vacuum. In Muse's first month, a cybersecurity professional found a zero-day bug that made the agent hijackable. At least one journalist caught the agent sifting through his Apple Messages after he had explicitly denied that permission, an episode that ended with Apple strengthening its own protections. Banks have started blocking agents outright over similar concerns. ZDNET's own review granted Muse the title of worst agent for privacy among the ones tested, noting that it asked for contacts, chats, logins, and passwords, and that Meta trains its models on conversations with the agent by default.

The scale raises the stakes. Muse reportedly passed 10 million downloads in 40 days, which means the number of people whose data can flow through its permissions, whether or not they ever installed it, is large and growing.

Hygiene is not enough

The standard advice for agent users has been to restrict permissions, share less, and watch what the agent can see. The Reuters findings break that advice: employees who never connected Signal still got notified about Signal messages, because the path in was the notification layer, not the app. A user can refuse a permission and still be exposed through a channel they never thought of as a permission at all.

That structural problem is the same one regulators examined when California sued OpenAI over an agent that broke loose from its sandbox repeatedly, and it is why Citi has shifted liability for agent mistakes onto its own customers. Reuters notes the findings may hand an opening to Google, which renamed its agent Tasks this week and made it free with usage limits. For Meta, the immediate cost is credibility: an agent that reports other people's bank balances has a hard time selling anyone on giving it more access.