The post went up on Meta’s internal forum this fall, one of more than 20 that Reuters reviewed while reporting on the company’s artificial intelligence assistant. An employee wrote that the chatbot, called Muse, had handed over a bank account balance that belonged to a different person. It was not an isolated complaint on the thread.
The details are mundane in the way that privacy failures usually are. Muse surfaced a balance of $2,722.58 held at Capital One, and another of $5,899.57 at Charles Schwab, according to the internal posts. It also offered up parenting advice nobody asked for, career information about strangers, and details of orders placed by other people.
Why this deserves a few minutes of an ordinary reader’s attention: the assistant is built to read across the apps already on your phone. Most people keep a bank app, a family calendar, a work mailbox and a messaging service on one device, and the value of an assistant comes precisely from crossing those lines. When it crosses them in the wrong direction, the convenience becomes the leak.
Muse launched in September and has been downloaded more than 10 million times, which puts it among the most widely distributed consumer AI products in the world. It runs inside the apps Meta already owns, including WhatsApp, Instagram and Messenger, and can act on what it finds there.
The mechanism matters more than the anecdotes. One employee told the forum they had never granted Muse access to other applications. A colleague pointed to a simpler explanation: on Android, the assistant can read push notifications by default. A push notification is the banner that slides down from the top of a phone screen when a bank, a delivery service or a colleague sends a message. That banner often carries the message itself, including the amount on a deposit or the text of an encrypted chat.
An app that can read every banner sees a great deal. It sees that your card was charged, that a package shipped, that a relative sent a message on Signal. The assistant does not need to break into the banking app to learn a balance. It only needs to read what the phone already displays.
Android treats that access as a single switch rather than a menu. A user who grants it hands over the whole stream, not a chosen list of senders, and the setting lives in a permissions screen most people never open. There is no version of the arrangement in which the assistant sees delivery updates but not bank alerts.
The industry has met this problem before. Microsoft delayed and then redesigned its Recall feature, which was meant to store screenshots of everything a user saw on a Windows machine, after researchers showed how easily the archive could be searched by anyone with access to the device. The lesson from that retreat was not that the feature was impossible, but that a product which records everything has to answer for everything it records.
Perhaps the strangest example in the internal posts involved Signal, the encrypted messaging service. Muse reminded a user to reply to a message that had arrived on someone else’s phone. The content had travelled from an encrypted app into a general-purpose assistant, through a channel designed to be readable by the operating system rather than by the user’s other software.
Meta’s position is that these are isolated incidents. The company said it takes the reports seriously and will keep updating the protections around the product. It has not published a full account of how many users were affected or what kinds of data were exposed.
There is a fair version of the counterargument, and it does not require defending the company. Assistants this new will misfire, and the first generation of any software that touches a phone’s notification stream tends to discover the boundaries by crossing them. The failures described in the internal forum are the kind a careful engineering team can patch, and the willingness of employees to post about them internally is itself a sign that someone is watching.
The harder question is whether the permission was ever explained. Reading notifications is a choice a platform makes on a user’s behalf, and the person holding the phone is rarely asked whether a bank alert should be visible to an assistant that may summarize it to someone else. The default is doing the work that consent should be doing.
Regulators have started asking similar questions about how AI products handle data drawn from other applications, and the answers will shape how quickly assistants move from answering questions to acting on a user’s behalf. A product that sends a message, books a table or moves money needs a firmer boundary than one that drafts a caption.
For the employee who wrote the original post, the boundary was already crossed. The assistant had their bank balance confused with somebody else’s, and no setting on their phone had warned them it could happen. The thread they started is now the clearest public record of what that looks like from the inside.


