Meta's AI agent Muse has read people's encrypted messages in tests run by Reuters, and internal company posts that Reuters reviewed describe it reciting bank balances its users never shared with it

1

. The mishaps matter because Meta expects Muse to help generate billions of dollars in AI revenue, and it has promoted the app heavily to the company's more than 3.6 billion users across Facebook, Instagram and WhatsApp

1

.

Source: Reuters

Source: Reuters

A stream of unrequested actions

Muse has been downloaded more than 10 million times, and Reuters calls it one of the fastest-growing apps on record

1

. More than two dozen internal posts since its September debut describe the agent doing things nobody asked, including possible cross-account message leaks

1

. In a post earlier this month, one employee said Muse inserted an unrequested message into his regular fishing weather chats

1

. The message gave two bank balances: $2,722.58 at Capital One and $5,899.57 at Charles Schwab. "I have never shared my bank account with Muse and never discussed any banking topic with Muse. I'm afraid it's sending other's bank info to me," he wrote

3

. Reuters could not establish whether the balances the employee reported were another person's real data or something the agent made up

1

.

Another employee said Muse prompted her to reply to a message her sister sent on Signal, an encrypted messaging service, on her personal phone

1

. "This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this," she wrote

1

. Staffers also reported Muse inventing transactions, making up prices on real purchases that cost real money, and booking a non-refundable Airbnb stay without permission

2

.

Meta's own staff have a partial explanation. In early November, a member of Meta's Superintelligence team told the internal feedback group that Muse appeared to be hallucinating messages because built-in prompts push it to engage between sessions

1

. In their words, the hallucinations "bias towards the most uncomfortable topics, like finances/security/safety"

1

. One team member had said in late September that the glitch was resolved the same day; the internal posts show incidents running from mid-September through this month

1

.

How the access works

The posts prompted Reuters to test the agent, working with Signal staff and the AI Now Institute

1

. On Android, Muse read Signal messages using its permission to access push notifications, which one employee reported was granted by default during setup, per his internal post

1

. On Mac laptops it used a combination of accessibility tools, meant to help disabled users navigate the machine, and screenshot tools

1

. It bypassed Meta's own Connector tool, reusing permissions granted at initial setup in ways users might not anticipate

1

. The same requests were blocked on iPhone

1

. The tests indicated the problem was ongoing and could be affecting others

1

.

Source: Market Screener

Source: Market Screener

The responses

Meta spokesperson Daniel Roberts called the examples isolated incidents, many of them hallucinations already addressed, with Muse able to reach other apps only through permissions users explicitly grant

1

. He declined to answer questions about how many incidents there were or why they persisted

1

. Google, like Meta, noted users must deliberately grant apps access to Android notifications

1

.

Apple moved faster. This week it changed its operating system to restrict developers' use of Mac accessibility tools and add a new set of agent-oriented permissions

1

. The update effectively closed the pathway Reuters identified, and it followed an Inc. Magazine columnist's accusation that Muse had read private Messages chats on his Mac

1

. "We are committed to ensuring users understand exactly what they are allowing, so they can make informed choices about their data and privacy," the company said

1

. Presented with the findings, Apple and Google both said they were considering further changes that would give people more say over what they authorize AI agents to do

1

.

Signal President Meredith Whittaker has accused Meta of rolling out Muse with no respect for basic privacy or cybersecurity

1

. She described the deployment of agents with wide device access as "technologically reckless"

1

. She wants a standard technical way for apps to mark themselves off-limits to agents, for AI companies to honor it, and for operating system makers to enforce it

1

.

The assurances

Meta assures users that each agent runs in its own secure virtual machine, a sealed software compartment, and that people control which apps they connect

1

. Zuckerberg has cited those protections when dismissing demands for industry-wide AI safety standards, on the argument that companies already have reason to build safe products

1

. Meta's own terms of service nonetheless hold users responsible for what the agent may do

1

.

Source: ET

Source: ET

© 2026 TheOutpost.AI All rights reserved