Technology•

Meta's Own Testers Called Muse AI's Access to Encrypted Signal Messages an 'Extreme Privacy Violation'

Internal Meta chats show employees warning that the Muse AI agent read encrypted Signal messages and mixed up user conversations. Independent testing confirmed two access routes; Meta blames isolated errors.

Written By :MINRK
Meta's Own Testers Called Muse AI's Access to Encrypted Signal Messages an 'Extreme Privacy Violation'

Photo by NordWood Themes on Unsplash

1. Meta Muse AI Draws Privacy Warnings From Inside the Company

Meta employees testing the company's Muse AI agent warned internally that it appeared to read encrypted private messages, including Signal conversations, and could mix up data between users, according to internal chats detailed in an investigation published on November 18, 2026. One staff member described the access to Signal messages as an "extreme privacy violation."

Muse, which Meta launched in September 2026, is an AI agent: software designed to carry out tasks on a person's behalf across the apps and services on their devices. The concerns raised by staff echo complaints from early adopters, who said the agent seemed to be looking at their private messages and behaved in ways they could not predict.

Signal is a messaging service built around end-to-end encryption, a design in which only the sender and the recipient can read what is sent. People choose it precisely because they expect that no third party, human or automated, will see their conversations. A report that an AI agent from one of the world's largest technology companies could read those messages therefore touches a particularly sensitive category of personal data.

The significance of the internal chats is that they show the same worries being voiced by the people responsible for testing the product. The problems were not only reported from outside by users; Meta's own workforce had flagged them in the company's internal channels.

2. What Employees Said Muse Was Doing

Three categories of behavior stand out in the internal communications. The first is access to content that users would not expect a bot to review. Staff testing Muse warned that the agent appeared to be reaching encrypted private messages, and it was this behavior, as it applied to Signal, that one employee labeled an extreme privacy violation.

The second is the handling of information belonging to different people. Employees said Muse could confuse conversations between users, raising the possibility that sensitive details belonging to one person might be shown to another. For a product that works with personal data, a failure to keep accounts separate is among the most serious faults that can occur.

The third involves messages that nobody requested. Staff reported that Muse sent notifications to users without their consent. Among the examples cited were notifications presenting bank balances and personal reminders.

Unprompted notifications matter for a reason beyond annoyance. An agent that volunteers financial details or reminders is revealing something about the information it has gathered and the conclusions it has drawn from it. When that happens without the user having asked, it becomes hard for the person to know what the software has seen, where the information came from and whether it is accurate.

Each of these issues would be notable by itself. Together they describe an assistant that reached further into private data than expected, did not always keep one person's information apart from another's, and communicated with users in ways they had not asked for.

3. How Did Muse Reach Encrypted Messages?

Muse reached private messages through two routes: broad push-notification permissions on Android and accessibility features on macOS. Neither route was a deliberate integration with messaging apps. The finding came from testing carried out with the help of Signal staff and the AI Now Institute, a research organization, which means the conclusion did not rest on the internal chats alone.

Both routes rely on features that exist for ordinary reasons. On a phone, an incoming message is often displayed as a notification that includes some of its text. An app that has been granted wide access to notifications can see that text as it appears. On a computer, accessibility features are built to let assistive software read and interact with what is shown on screen, and an app given those permissions can observe the content of other applications.

This explains how an encrypted service can be affected. End-to-end encryption protects a message while it travels between devices, so that it cannot be read in transit. Once the message arrives and is displayed to its recipient, it exists in readable form on that device. Software with sufficiently broad permissions on the device can see it at that point, without breaking the encryption at all.

The distinction is central to understanding the findings. Nothing in them indicates that Signal's encryption was defeated or that the messaging service itself was at fault. The exposure occurred at the edge, on the user's own device, after the protected message had been delivered and decrypted for display. From the user's point of view, however, the result is the same: content they believed was private was available to software they had not knowingly authorized to read it.

The outcome differed on Apple's phones. The same pathway was blocked on the iPhone, so the access that worked on Android and macOS did not succeed there.

4. Meta Attributes the Incidents to Isolated Errors

Meta has rejected the idea that the episodes reveal a fundamental defect. The company attributed the incidents mainly to isolated hallucinations and to device permissions that were configured incorrectly. It said the problems do not amount to a systematic flaw in the way Muse handles data, and that many of the issues have already been fixed.

A hallucination, in the context of AI models, is output that the system produces with apparent confidence but that is not grounded in real information. Describing an incident this way implies that the agent generated incorrect content, as opposed to retrieving and exposing genuine data it should not have had.

The company's account sits uneasily beside the results of the independent testing. That testing, supported by Signal staff and outside researchers, identified specific technical routes by which Muse obtained message content. A route that works reliably through a granted permission is a different kind of problem from an occasional erroneous output, because it concerns what the agent is able to see, not what it mistakenly says.

Meta's position is further complicated by the distance between what staff were saying internally and how the product was behaving in public. When a company's own testers use language as strong as an extreme privacy violation, the concern is difficult to treat as routine feedback of the kind that is logged and resolved quietly.

5. Why AI Agents Raise a Different Permission Problem

The dispute over Muse illustrates a wider challenge for the technology. Earlier AI assistants mostly answered questions inside a single app. Agents are meant to take actions for the user, and to do that well they need context: what the person is working on, what messages they have received, what is on their screen.

That goal creates a strong incentive to collect as much surrounding information as the operating system will allow. The permission systems on phones and computers were designed for apps with narrower purposes, and a single broad grant, such as access to notifications or to accessibility features, can open far more data than a person installing an assistant would reasonably anticipate.

Consent is the weak point in that arrangement. A person may approve a permission request during setup without appreciating that it lets the agent see the text of every incoming notification, including those from apps chosen specifically for their privacy. The approval is technically valid, yet the user's understanding of what was granted can be far narrower than what the software is then able to do.

For that reason, the technique at issue is unlikely to be unique to Meta. Notification permissions and accessibility interfaces are available to any developer, and any agent that aims to be useful across an entire device will face the same pull toward ambient data. The findings suggest that agents require a different approach to permissions, one that tells users plainly what an agent can read and limits it to what a task needs.

The difference between platforms in this case also shows how much depends on the operating system. Where the route was closed, as on the iPhone, the access did not occur. Where the permissions were broad, as on Android and macOS, it did.

6. What Is at Stake for Meta

The timing is awkward for the company. Meta has tied a large part of its near-term revenue story to earning money from artificial intelligence at scale, and Muse is central to that effort. A product meant to demonstrate the commercial promise of AI agents is instead facing questions about what it reads.

Agents depend on trust to an unusual degree. A person who allows software to act for them, across their messages and other apps, has to believe it will stay within the limits they expect. Reports that an agent read conversations in an encrypted messenger, or showed one user material connected to another, go directly to that belief.

Trust of this kind is slow to build and quick to lose. An assistant that answers a question badly can be corrected or ignored, but an agent that is thought to have looked where it should not have is likely to be switched off, and the permissions it relies on withdrawn. For a product whose usefulness depends on those permissions, that is a commercial risk as well as a reputational one.

The credibility of the findings adds to the difficulty. The concerns come partly from Meta's own internal communications and were supported by independent testers, including staff from the messaging service whose content was accessed. That combination is harder to dismiss than complaints from individual users.

Meta says that many of the issues have been addressed. The open questions are whether the broad permissions that made the access possible will be narrowed, and whether users will be told clearly what Muse can see on each platform. How the company answers them will shape how its agent, and competing products built on similar permissions, are received.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY