Infographic showing AI assistant permissions, operating-system privacy controls, and encrypted messaging on Android and macOS devices.
Meta's Muse AI agent read Signal message text on an Android phone and a Mac without Signal ever being connected to the assistant. This is not a break in Signal's encryption. It is what happens when an AI agent holds broad operating-system permissions, and that makes it an endpoint-management problem for anyone running mixed fleets.

Infographic showing AI assistant permissions, operating-system privacy controls, and encrypted messaging on Android and macOS devices. What Reuters found​

Reuters worked with Signal and the AI Now Institute to test whether Muse could retrieve private messages across different operating systems. The tests followed complaints from Meta employees. Reuters reviewed more than two dozen internal Meta posts flagging Muse problems between mid-September and November.

The results by platform:

  • Android: Muse pulled Signal message text using notification permissions granted during setup, without going through Meta's own tool for linking apps. One internal poster reportedly found that Muse had permission by default to read push notifications on Android phones.
  • Mac: Muse used Apple's accessibility and screenshot tools to read conversations displayed on screen.
  • iPhone: The same access attempts were blocked on an iPhone. That is a result of these specific tests, not a guarantee for every iOS configuration.

Why encryption wasn't the issue​

Signal's end-to-end encryption protects messages in transit. Once a message is decrypted and shown as a notification preview or in an app window, any software with sufficient device permissions can see it. As eSecurity Planet put it, Muse could read private Signal conversations even when the messaging app was not connected to the assistant.

The mechanism is ordinary. Android's own documentation describes notification listeners as services that receive callbacks when notifications are posted or removed. A user who grants one to an assistant is, in effect, giving it a live feed of everything that pops up. Mac accessibility and screen-capture rights work the same way for anything visible on screen.

The important distinction is between a service-level connector (the app is linked to the assistant) and device-level access (the assistant can see what the OS shows). Meta's connector controls only govern the first.

Meta's response and what remains unproven​

Meta disputes the wider picture. Spokesman Daniel Roberts told Reuters that Muse can reach other applications only through permissions people explicitly allow. Meta characterised the incidents as isolated rather than evidence that every Muse user was affected. It also attributes many suspected account mix-ups to inaccurate AI-generated details.

Some points should stay separate:

  • The internal complaints included unsolicited chats and financial details tied to unfamiliar accounts. Per eSecurity Planet, investigators could not verify whether the assistant exposed another person's information or generated inaccurate details. These should not be treated as confirmed data breaches.
  • Reuters' tests relied on device permissions outside the connector system. They did not establish that Meta's Sentinel access-control component was bypassed.
  • Muse's separate October scrutiny involved publicly discoverable shared-conversation links. That is a different mechanism from this device-permission issue.

Apple's macOS changes: conflicting accounts​

Apple's developer news post of November 16 says macOS 27.3 beta 2 splits Device Control and Data Access into two settings, System & App Protection Override and Accessibility. It also adds app selection from a list, an "ask every time" option and a transparency indicator when one app controls another. Apple says all previously granted permissions remain in effect — more precisely, its post states that existing grants stay active.

Reports differ on whether this closes the Muse route:

  • eSecurity Planet says the changes have not been confirmed to prevent the behavior shown in the Signal tests.
  • The Neuron and The Next Web, citing Reuters, say an Apple update this week closed that route on Macs.

I could not reconcile these from the evidence available. Because it is a beta, and because older grants persist, admins should not assume their Macs are protected. Test instead.

What IT admins should do​

The advice below combines eSecurity Planet's checks with Apple's and Google's documentation:

  1. Inventory permissions beyond connectors. Find AI assistants holding Android notification access, or Mac Accessibility and screen-recording rights. Remove any that an approved business use doesn't need.
  2. Check Android work-profile boundaries. Google describes Work Profile as separating work apps and data from personal ones, but its documentation does not promise that this blocks every listener route. Confirm which notification policies apply to your Android version, enrollment mode and management setup, and test them on real devices.
  3. Re-audit existing Mac approvals. New controls apply to new requests. Review grants already in place.
  4. Test with fictional messages. In a controlled setup, send dummy messages in an unconnected app and see what the agent can retrieve.
  5. Reduce notification previews. One consumer-focused guide suggests setting Signal's Android notification content to hide names and message text. That limits what a listener sees, though it is a mitigation rather than a fix.

Why this matters for Windows shops​

Mixed fleets are the norm, and agent-style assistants are arriving on every platform. The Muse case shows a pattern that applies equally to Windows desktops: an "isolated" app is only isolated from the assistant's connectors, not from the OS surfaces the assistant is allowed to observe. eSecurity Planet also notes that Signal enabled screenshot protection by default on Windows 11 in response to Microsoft's Recall feature, though that safeguard did not extend to Macs. I could not independently verify that detail's date and configuration, so treat it as eSecurity Planet's account.

The practical lesson is to treat AI-agent approval as a permissions review, not a connector review. Ask what the agent can see, not just what it is linked to.

 

References

  1. Meta Muse AI Read Signal Chats Through Android and Mac Permissions - eSecurity Planet eSecurity Planet 2026-11-20T16:00:00+00:00
  2. What is an Android Work Profile? - Android Enterprise Help support.google.com
  3. Meta's Muse AI Can Read Your Encrypted Signal Messages, Tests Show allaboutcookies.org