Meta's new AI agent Muse, in tests conducted by Reuters, accessed end-to-end encrypted messages in Signal on Android using only notification access permission, and on Mac using accessibility permission. Meta described this as an "isolated incident." But the problem isn't malicious behavior by the agent; it's that mobile operating systems' permission architecture was never designed with AI agents in mind.
What happened?
According to Reuters' testing, Muse was able to see the content of messages from Signal on Android when granted notification-reading permission; on macOS, it achieved the same result by reading on-screen text via accessibility permission. Signal uses end-to-end encryption, meaning messages travel in a form that cannot be read on the server or over the network. Muse didn't pull the message from the server — it read text that had already been decrypted on the device and was visible on screen. Meta characterized the incident as a one-time error and did not acknowledge it as a systemic design issue.
Why does it matter?
Notification access and accessibility permissions have existed for years; the former was designed so smartwatches could mirror notifications, the latter so screen readers could narrate the interface to visually impaired users. Neither was written with the scenario in mind of an AI agent with memory and reasoning capability collecting and interpreting this data. Spyware like Pegasus abusing the accessibility service, or certain ad SDKs reading notifications for analytics purposes, was long considered "abuse." Now the same behavior shows up as a feature approved on a permission screen, in the official product of a major tech company.
This isn't the first time this debate has centered on Signal. Microsoft's Recall feature, introduced in 2024, drew heavy backlash because it periodically saved everything the user saw on screen as an image; Signal's own executives openly said at the time that Recall rendered end-to-end encryption meaningless. The Muse case is a different application of the same principle: the agent isn't breaking the protocol — it's looking at the layer the protocol can never protect, the screen and the notification interface.
The critical point here: encryption protects the data path and storage, not the text displayed on screen. Any software running on a device with access to the visual or textual interface sees what the user sees, no matter how strongly the message was encrypted. This isn't a flaw in Signal's protocol; it's the natural limit of the guarantee encryption can provide on an open-ended operating system.
Who does this affect, and how?
The problem isn't unique to Signal; any messaging app that uses the same notification and accessibility permissions — WhatsApp, Telegram, iMessage — is theoretically exposed in the same way.
| Permission type | Original design purpose | How it can be used by an agent |
|---|---|---|
| Notification access | Smartwatch sync, notification management | Bulk reading of message previews |
| Accessibility | Screen readers, access for users with motor impairments | Reading any on-screen text, app-independent |
The group most at risk is users who treat the "end-to-end encrypted" label as an absolute security guarantee. On the enterprise side, teams using Signal for compliance reasons need to re-audit which apps on their devices have been granted notification or accessibility permissions. For Meta, the real risk is reputational: given the company's history with Cambridge Analytica and WhatsApp data sharing, the "isolated incident" explanation will continue to be met with skepticism.
What's next?
There are three things worth watching. First, whether Google and Apple will restrict notification and accessibility permissions with a dedicated approval layer specific to AI agents. Second, whether Apple Intelligence and Google's own agents pass the same test; if the issue is specific to Muse, it's a minor story — if it's platform-wide, it's a much bigger one. Third, whether Signal will make client-side changes such as limiting notification previews by default or adding extra protection against accessibility services.
Meta's explanation isn't technically wrong — Muse wasn't designed with malicious intent as spyware. But the "isolated incident" framing overlooks the real issue: it's not the agent's behavior that's at fault, but the fact that the permission system was never adapted for the age of agents.
Yorumlar (0)
Henüz yorum yok. İlk yorumu sen yaz.
Yorum yapmak için Sinyal'i indir
Yorumlar Sinyal hesabıyla yapılır. Mobil uygulamada giriş yap, yorum bırak.