Meta's AI Agent Read Private Signal Messages. How Did It Get Access?

Meta employees reported strange messages from Muse, while Reuters confirmed the AI agent could retrieve private Signal messages through existing device permissions. The findings raise questions about how much access users are granting AI assistants.

Written By
Marianne Sison
Marianne Sison
Nov 20, 2026
5 minute read

A Meta employee asked the company's AI assistant about fishing conditions. Somehow, the conversation ended up with bank balances from Capital One and Charles Schwab.

Here's the strange part: The employee said neither account belonged to them. They hadn't even connected their bank accounts to the assistant.

And apparently, other Meta employees were having similarly bizarre experiences.

According to a Reuters investigation, Meta's AI agent Muse has been sending employees messages about unfamiliar purchases and conversations they never started.

Meta says many of these incidents were AI hallucinations. But Reuters also confirmed something more concerning: Muse could access private Signal messages through existing device permissions, even when users hadn't connected Signal directly to the assistant.

So, how much of your personal information can an AI agent access once you've given it permission to help?

Meta employees are getting messages they never asked for

Advertisement

Meta introduced Muse in September as an AI assistant that could handle everyday tasks, including scheduling appointments and making purchases.

The idea sounds appealing. You give the assistant a task, and it takes care of the work across your apps. But some employees testing Muse started receiving messages that didn't make sense.

Reuters reviewed more than two dozen internal posts describing unusual behavior. One employee received a message about picking up a coffee table, even though they hadn't discussed buying furniture.

Another said Muse booked a non-refundable Airbnb stay that they never authorized.

Then there was the employee who received those unexpected bank balances during a conversation about fishing.

Naturally, some employees wondered whether Muse was accidentally pulling information from other people's accounts.

Meta has a different explanation.

The company said many of the suspicious messages were hallucinations. In other words, Muse may have invented conversations or financial details that sounded real.

Reuters couldn't confirm whether the bank balances actually belonged to another person.

Advertisement

And there's an important difference here. An AI assistant inventing a bank balance is a reliability issue. An assistant retrieving someone's financial information would be a privacy breach.

Meta also told TechCrunch that employees sometimes test experimental versions of Muse, so their experiences might differ from what public users encounter.

The Signal incident is where things get more concerning

One employee reported that Muse unexpectedly reminded her to reply to a message from her sister on Signal.

She hadn't connected Signal to the assistant.

Signal is known for its end-to-end encryption, which protects messages from being read by outsiders while they're transmitted between devices.

So how did Muse know about the message?

Reuters worked with Signal and the AI Now Institute to investigate. Their tests found that Muse could access message content through permissions already available on the device.

On Android, the assistant could retrieve Signal messages through notification access.

On macOS, it could access information through accessibility features and screenshots.

The researchers couldn't reproduce the same behavior during their iPhone tests.

Advertisement

To be clear, Muse didn't break Signal's encryption. Once a message reaches your device, its contents can appear in notifications or on your screen. Apps with certain permissions may be able to read that information.

Here's where things get uncomfortable.

You might give an AI assistant permission to check notifications because you want help managing appointments. But the same permission could expose previews of private conversations.

And if the assistant can decide when to retrieve information, it might access messages that have nothing to do with your original request.

You gave the AI permission. But did you know what you agreed to?

Meta says Muse requires users to grant permission before it can access other applications.

Technically, that's how many AI assistants work. They ask for access to your device or connected apps so they can complete tasks on your behalf.

But the Reuters investigation shows why those permissions deserve a closer look.

Consider accessibility access on a Mac. The feature helps software interact with on-screen content. An AI assistant might use it to navigate applications or complete repetitive tasks.

Advertisement

However, accessibility permissions can also expose information from applications that the user never intended to connect.

A similar concern surfaced in September when journalist Jason Aten reported that Muse accessed his private Mac messages despite his decision to deny access.

Meta disputed his account and said the Messages integration required explicit authorization. Aten maintained that the agent accessed his messages anyway.

The disagreement remains unresolved.

What makes AI agents different from ordinary apps is their ability to decide which actions to take. You might ask an assistant to organize your schedule. During that task, the assistant could determine that checking your messages would help.

Would you have approved that decision if it had asked first?

The problem is that permission screens don't necessarily explain everything an autonomous assistant might do with the access you've granted.

Advertisement

Apple has already made changes to limit AI access

The Muse investigation also caught the attention of operating-system developers.

According to Reuters, Apple updated macOS during the week of November 16 to restrict certain accessibility-based access methods and introduce permissions specifically for AI agents.

The changes blocked the method that researchers had used to retrieve Signal content on a Mac.

Google, meanwhile, said Android users must deliberately approve notification permissions. The company was also considering additional protections.

Signal President Meredith Whittaker wants app developers to have more authority over whether AI agents can access their content.

Her concern makes sense. Someone might want an assistant to summarize work notifications while keeping private conversations off-limits.

Right now, broad device permissions can make those preferences difficult to enforce.

More specific permissions could allow someone to authorize an assistant to retrieve a calendar invitation while blocking access to private messages.

Advertisement

But companies would need to build those restrictions into the software itself.

Who's responsible when your AI assistant goes too far?

The concerns become even more complicated as AI agents start taking actions instead of simply answering questions.

Now consider what happens when an assistant has permission to book travel or manage your finances. If it makes a purchase you didn't approve, who is responsible? And how do you prove which instructions you actually gave?

Companies can require confirmation before sensitive transactions. They can also maintain records of the assistant's activity, including completed purchases and user approvals.

Reading private messages presents another challenge.

By the time an assistant asks whether it should share or use the information, it may have already accessed something you wanted to keep private.

For anyone using an AI agent, it's worth reviewing notification permissions and accessibility access, especially if the assistant can operate across multiple applications.

Advertisement

Disconnecting an app through the assistant's settings may not revoke permissions granted through the operating system.

The bigger question is how AI companies will handle those permissions as assistants become more independent.

Muse's strange bank updates might turn out to be hallucinations. But the Signal tests confirmed that an AI agent could retrieve private information through permissions intended for other purposes.

And as these assistants take on more responsibility, users need to know exactly what they're authorizing.

If you ask an AI assistant to help manage your day, how much of your private life should it be allowed to see?

Marianne Sison

Marianne is a technology analyst with nearly five years of experience reviewing collaborative work management solutions. She helps businesses identify the right tools and apply best practices to streamline workflows and improve project performance. Her insights on project management and unified communications appear in publications like Project-management.com, TechRepublic, and Fit Small Business.

The Neuron Logo

Don't fall behind on AI. Get the AI trends & tools you need to know. Join 700,000+ professionals from top companies like Microsoft, Apple, Salesforce and more.

Property of TechnologyAdvice. © 2026 TechnologyAdvice. All Rights Reserved

Advertiser Disclosure: Some of the products that appear on this site are from companies from which TechnologyAdvice receives compensation. This compensation may impact how and where products appear on this site including, for example, the order in which they appear. TechnologyAdvice does not include all companies or all types of products available in the marketplace.

Stay in the loop

Get notified when we publish new articles.