Reporting based on Reuters tests, relayed on November 18-19, 2026, says Meta's Muse AI agent read Signal message text on Android phones and Macs without ever connecting to Signal. The tests did not show that Muse broke Signal's encryption. They showed something more mundane and more useful to understand: an agent that holds operating-system permissions can read whatever the operating system already shows it. If you give any AI agent notification access, accessibility or screen capture, the question "which apps did I connect?" is the wrong one. The right one is "what can the OS let this agent see?"
We could not open the Reuters piece itself, so this post relies on Cyber Security News's summary of it and on Decrypt's earlier reporting on a related dispute over Muse and macOS Messages. Treat every figure below as reported, not independently reproduced by Beam.
| Question | Short answer |
|---|---|
| Am I affected? | Only if you run Muse (or another agent) with Android notification access, or macOS accessibility or screen recording. No version or patch is named in the coverage |
| Is there a fix? | No patch is described. The fix is revoking OS permissions and hiding notification text |
| Was encryption broken? | No. The tests did not show that |
| Does data leave my machine? | Unknown from the coverage. Whether Muse sends what it reads to Meta servers is not stated in the sources we read |
| Is it confirmed? | Partly. Reuters' tests are reported; Meta calls the cases isolated and some reports hallucinations |
What Reuters reportedly found
Per the summary, Muse read Signal text on both platforms, with no Signal "Connector" added in Meta's own app:
- Android: Muse read message text from push notifications. A notification preview contains the message text once it is decrypted and displayed, and any app with notification access can read it.
- macOS: testers found Muse could use accessibility and screenshot tools to read Signal's window.
- iPhone: the same request was blocked on the iPhone used in testing.
- Revocation gap: leaving Signal off the agent's list of connected apps does not revoke permissions granted through the operating system.
The background reportedly began with an internal complaint: a Meta employee said Muse reminded her to reply to a Signal message from her sister, on a personal phone, although she did not recall granting it access to other apps. A colleague traced similar reminders to Android notification access. Employees also reported unsolicited messages and one update that included two bank balances the recipient did not recognize. The tests confirmed Muse could read Signal text but did not establish that the bank figures belonged to another user.
Why this is not an encryption break
Signal's end-to-end encryption protects messages in transit and at rest on Signal's servers. Once the app on your device decrypts a message, it must show it to you. Anything the OS lets another process observe at that stage is outside the cryptography. Android notifications can carry the plaintext preview. A macOS accessibility client can read the text of UI elements. A screen-recording permission captures pixels. None of this needs to touch Signal's code.
This is why the revocation gap matters. Muse has two layers of consent: its own connector list (what Meta's app says it is connected to) and the OS grants (what the system lets it observe). Users reasonably think the first controls the second. According to the tests, it does not.
How the Android and Mac paths differ
The two platforms leak for different reasons, and the fixes differ too.
Android notification access is a single, coarse permission. An app that holds it receives every notification posted by every app, including title and body text, and can read them as they arrive. It is designed for smartwatches, automation tools and notification managers. An AI agent that asks for it to "remind you about messages" gets the same feed. The only app-side control is what the sending app puts in the notification. Signal lets you hide the content, which turns the preview into a generic alert. If previews are hidden, a notification reader gets nothing to read.
macOS accessibility and screen recording are broader. Accessibility lets a client inspect the interface tree of other apps, including the text of messages in a window. Screen recording captures what is drawn. Both read the rendered result, so they work whether or not the app has any "connector" with the agent. They also persist: once granted to a process, they stay granted until you remove them in System Settings.
The iPhone result is worth noting. The same request was reportedly blocked on iOS, where there is no equivalent of a global notification reader or accessibility client for third-party apps. This is a platform design difference, not a statement about Meta's agent.
The practical lesson: the strength of an agent's consent screen tells you little. The OS permission list tells you what is possible.
The Meta disagreement, and an older dispute
Meta's position, as summarized: Muse accesses other apps only through permissions users explicitly grant; the reported cases are isolated; many suspected account mix-ups were AI hallucinations. Signal's own response was not in the coverage we read.
There is also an earlier, separate dispute reported by Decrypt in the fall. Inc. columnist Jason Aten said he declined Messages access during setup, confirmed Full Disk Access was off, and Muse still used a private text exchange in a column pitch; asked how it knew, Muse said it had only seen notification banners. Meta's head of communications Andy Stone said on X that the Mac Messages integration is opt-in and needs two permissions (macOS Full Disk Access and Muse's own Messages connector). Meta's David Singleton said the banner explanation was wrong and that Muse syncs Messages data only when access is enabled. Aten's claim that Muse synced his local Messages database is his report, not verified by Meta. The detail worth keeping: both disputes involve the gap between what an agent says it can see and what the OS allows it to see.
| Claim | Source | Status |
|---|---|---|
| Muse read Signal text on Android via notifications | Reuters tests, via Cyber Security News | Reported, not reproduced by us |
| Muse read Signal text on Mac via accessibility or screenshots | Same | Reported |
| Encryption was broken | Nobody | Not shown |
| Muse reads data without any user grant | Meta denies | Disputed |
| Bank balances came from another user | Employee report | Not established |
| Muse synced a user's Messages database with Full Disk Access off | Aten, via Decrypt | Disputed by Meta |
What this means if you run AI coding agents
Muse is a consumer agent, but your coding agent sits on the same OS and inherits the same grants. A terminal app with Screen Recording or Accessibility permission lets every agent launched from it use that permission. Concrete checks:
- List grants on macOS. System Settings, Privacy and Security: review Accessibility, Screen Recording, Full Disk Access, Input Monitoring and Automation. Remove entries for terminals, IDEs and agents you do not need to have them. Narrower AI permission controls are covered in our macOS 27.3 AI permission controls post.
- List grants on Android. Search settings for "Notification access" and "Accessibility" and remove agents that do not need them.
- Hide message text in notifications. In Signal's settings, switch notification content to hide the sender and message. You keep the alert, and any notification reader sees nothing useful.
- Treat connectors and OS grants as separate switches. Revoke both. Removing an app from an agent's connector list does not remove what the OS permits.
- Ask what leaves the device. The coverage does not say where Muse sends what it reads. For any agent, check the vendor's data documentation before granting screen or notification access.
- Separate profiles. Run agents in a user account or device that does not hold your private messaging, banking or work chats.
For background on agent permission hygiene, see How to keep Muse, Dots, Claude Duet and Grok Bot safe and the earlier Meta Project OT post on unchecked coding agents. For the general approach to scoping what an agent may touch, read Beam's introduction.
What Beam can and cannot do here
Beam watches coding agents such as Claude Code and Codex through hooks. It can flag risky tool calls (credential exposure, destructive commands, privilege escalation), redact secrets before storing events, and keep data on-device, as described on the local-first section and what Beam watches. Its deny is hook-based for Claude Code and Codex PreToolUse, not a kernel sandbox, and policy is fail-open when no cached policy is available.
Beam does not monitor a consumer agent like Muse, and it cannot see or block what a different app reads through Android notification access or macOS accessibility. Those controls live in the operating system. Use the permission checks above for that layer.
What to watch next
Open questions: what data Muse transmits after reading a notification, whether Meta changes how connector settings relate to OS grants, and whether Signal responds. We will update this post if Meta or Signal publish details. Meta's own privacy and help pages are where an official answer would appear.
Frequently asked questions
Did Meta Muse break Signal's encryption?
No. According to the reporting, the tests did not show that Muse broke Signal's end-to-end encryption. The agent saw text already displayed on the device, through Android notifications or Mac accessibility and screenshot tools.
Is Meta Muse safe to use?
It depends on what you grant it. Meta says Muse can reach other apps only through permissions users give, but the reported tests show OS-level grants work even when an app is left off the agent's own connector list. Review notification access, accessibility and screen recording before relying on it.
How do I stop an AI agent reading my Signal messages?
On Android, turn off Notification access for the agent and hide message text in Signal notifications. On a Mac, remove the agent from Accessibility and Screen Recording in System Settings, Privacy and Security. Removing an app from the agent's connector list does not revoke these.
What is Meta Muse?
Muse is Meta's consumer AI agent, launched in September. Reporting says it passed 10 million downloads. It can act across apps on a phone or computer when users grant it permissions.
What does Meta say about the reports?
Meta says Muse can access other apps only through permissions users explicitly grant, calls the reported cases isolated, and says many suspected account mix-ups were AI hallucinations. Signal's response is not included in the coverage we read.
Does this affect coding agents like Claude Code or Codex?
Not by this report, which concerns a consumer agent. The mechanism does apply: any agent running on your machine inherits the OS permissions of the terminal or app that launched it. Check what those grants cover.
Related reading
- macOS 27.3 AI permission controls
- Keep Muse, Dots, Claude Duet and Grok Bot safe
- Meta Project OT and unchecked coding agents
Details reflect public reporting as of November 19, 2026 and may change as Meta, Signal or Reuters publish more.

