Executive Summary
Meta's AI assistant Muse read the body text of end-to-end encrypted Signal messages. Reuters re-created the access paths together with Signal staff and researchers at the AI Now Institute, and the result was published on November 18. Signal's encryption itself was never broken.
The path ran through operating system features built for people. On Android, Muse used a notification access permission to read the preview text that appears on screen when a message arrives. On macOS, it combined accessibility tools with screen capture to reach the same characters. Neither path went through Connector, the mechanism Meta built for linking outside services. Messages were read for users who had never linked Signal at all. The same attempt was blocked on iPhone, and Apple narrowed the macOS path in an operating system update this week.
Sections 1 through 4 follow what is written in the Reuters reporting and in the official statements from Meta and Signal. Section 5 is the Pebblous reading.
Key Figures
The first two numbers say how many devices this assistant already sits on. The last two say how wide the reporting looked, and which side has not moved yet.
Source: the Reuters reporting as carried by BNN Bloomberg and Dataconomy.
10M+
Muse downloads
Passed within two months of the September 2026 launch in the United States
~25%
Daily user retention
About one in four people who installed it opens it again every day
24+
Internal posts Reuters reviewed
Posts employees wrote about odd behavior from mid-September through November
0
Reported regulatory actions
No regulator has been reported to have acted; Apple and Google said only that they are reviewing tighter controls on agent permissions
Two Permissions Reached Messages the User Never Shared
It started on one Meta employee's personal phone. Muse pushed a notification suggesting a reply to a Signal message from the employee's sister. It was a message the employee had never shared with Muse. On an internal board, they wrote that this "seems like an extreme privacy violation."
Signal encrypts messages between one device and another. The content stays unreadable while it passes through servers, and unreadable if it is intercepted somewhere on the network. That part held here. The path Reuters re-created with Signal staff and AI Now Institute researchers did not undo the encryption. It picks up characters that have already been decrypted and drawn on the screen. The AI Now Institute, which took part in the test, is a nonprofit that studies the social impact of AI.
On Android, an arriving message puts part of its body text into the notification preview, the few lines that show on the lock screen or in the notification shade. An app granted notification access can read the text of every notification on the device, and Muse read the preview with that permission. On macOS, it reached the same characters by combining the accessibility tool that reads screen elements aloud with screen capture.
The core of this case sits in where the two paths did not go. Meta built a separate feature called Connector so that Muse can work with outside services, and the user attaches those services one at a time, a calendar or a mail account. Signal was never on that list. The messages were read anyway. Two broad permissions, granted once at setup, reached the contents of an app the user had never chosen.
Why the iPhone Blocked the Same Attempt
The same attempt did not work on iPhone. That is not because Meta was more careful on one platform. The two operating systems handle notifications differently to begin with. Android left the permission to read another app's notification text as something a user grants once and an app keeps using, and Apple never opened that position as widely to outside apps.
Results split inside one company as well. The path blocked on iPhone was open on the Mac. Apple has not kept this position narrow on every device it makes. A phone and a desktop computer stand on different permission models, and the same app reaches a different range depending on which machine it runs on.
Apple sharply restricted access to accessibility tools in a macOS update this week. It also added a system that makes AI agents request the permissions they use separately. The macOS path used in the test was narrowed by that update.
Two things are worth holding onto here. One is that this was not a break in cryptography but a case decided by what the operating system allows an app to do. The other is that it was therefore a gap a platform can close by changing a setting. One of the two paths was in fact narrowed in the same week the reporting ran.
From One Report to 24 Internal Posts
The reporting did not stop at one employee's account. In the internal feedback channel where employees raise problems, Reuters confirmed more than 24 posts describing unwanted behavior after the Muse launch. They run from mid-September through November.
Other cases in the same channel differ in kind from the message access, but they point the same way. There were reports of unrelated conversations getting mixed in, and a case where a nonrefundable accommodation booking went through without permission. One employee wrote that they had been signed up for a pet insurance plan costing $888 a month and that it would not cancel. In early November, an employee on Meta's superintelligence team wrote that Muse appeared to have made up messages while responding to prompts nudging it to resume a conversation.
The cases have one thing in common. An assistant that acts on your behalf on the device leaves traces when it acts wrongly, not only when it reads wrongly. The event at the center of this reporting is still the Signal message access. The rest reads more accurately as the background that keeps that one case from standing alone.
Meta and Signal Split on What the User Knew
Muse came out in the United States in September 2026. The description Meta attached at the time was that the product had been built with privacy and security in mind from the start.
Meta spokesperson Daniel Roberts framed these cases as "isolated incidents." Muse reaches other apps only through permissions the user explicitly grants, Roberts said, and a good share of the suspected account mix-ups were hallucinations that have already been fixed. Meta said its extensive internal testing continues.
Signal president Meredith Whittaker stood on the opposite side. Whittaker criticized Meta for shipping Muse without basic respect for privacy and cybersecurity, and called the act of deploying an agent with broad access to the device "technologically reckless."
The user did grant the permissions. What is in dispute is how far the user understood those permissions to reach. Apple and Google both said they are reviewing stronger user controls over AI agent permissions, without giving a schedule. No regulatory action has been reported so far.
Why Pebblous Is Watching This Permission
Notification previews and accessibility features are passages built for people. They exist so that someone with full hands can glance at a screen and catch what arrived, and so that someone who cannot see the screen can hear the words on it. Nothing has classified these features as data collection paths until now, because the reader on the other end was a person. The moment an agent sits on the device, the same passage turns into an input channel a machine uses.
Three months ago, writing about Meta's on-device model, Pebblous put it this way: once the agent lives on the device, what it reads on every request is the data sitting on that machine. That was said about a design that takes files, calendars and messages as input. This reporting adds one case where the range of that input runs wider than what the user picked.
So the question moves. Until now, checking a data boundary meant asking which data someone agreed to hand over. Once an agent is inside the device, the question to ask alongside it is where that agent is sitting and what it can see from there. The first question is answered by looking at a list of linked services, the second by looking at a list of operating system permissions. In this case, the two lists did not match.
For an organization bringing an agent inside, there are four places to check.
- Does the agent installed on work devices hold broad permissions such as notification reading, screen access or accessibility? Who approved those permissions, and when?
- Are the list of explicitly linked services and the list of operating system permissions managed separately? Has anyone put the two lists side by side?
- Do notification previews for the company messenger and mail show the body text as it is? Where does hiding previews alone shrink the exposed range?
- Is there a record that lets you check later what the agent read and what it did? Every case in this reporting surfaced only after a person noticed something odd and wrote it down.
Editor's Note: Drawing a data boundary has long been a question of storage and access rights. Which table a given person can query was the boundary, and that boundary could be written down in a document. An agent on the device pulls the boundary out in front of the screen. Characters drawn on a screen are not a table, so they do not appear in a permissions document, and the volume read that way is not smaller than a table's. What this case leaves behind is not an apology or a single patch, but a coordinate for where the boundary has to be drawn again.
References
Original Reporting
- 1.BNN Bloomberg. (2026-11-18). "As Meta bets big on Muse AI agent, early safety mishaps surface." Carries the original Reuters reporting.
Follow-Up and Cross-Checks
- 2.Tech Times. (2026-11-19). "Meta employees raise Muse AI privacy concerns, exposing security flaws and chat leaks."
- 3.Dataconomy. (2026-11-19). "Meta's Muse AI agent accessed encrypted messages, Reuters reports."
- 4.CybersecurityNews. (2026-11). "Muse reads encrypted messages."
- 5.Martin Cid Magazine. (2026-11-19). "Meta Muse AI agent, Signal messages and Android notifications." Detail on the iPhone block and the Apple patch.