Signal screen security: what it blocks

Signal screen security: what it blocks

Signal’s screen security switch has existed on Android for years and almost nobody enables it. It sits two screens into Privacy, it is off by default, and it does one thing. Now the app turns it on by itself.

On iOS 27, Signal restricts screenshots of chats. On Android, it watches for an AI assistant reading the conversation and enables screen security when it finds one, according to the announcement coverage. Signal called OS-level AI an existential threat. That sounds like positioning until you look at what the flag underneath the feature can and cannot do.

The flag blocks pixels. That is the whole mechanism, and it explains every limitation that follows.

What the toggle does at the compositor

Android implements screen security with one window flag: WindowManager.LayoutParams.FLAG_SECURE, value 0x2000. Setting it marks the window’s surface as non-capturable, and the compositor leaves that layer out of the buffer it hands to capture clients. adb shell screencap returns black. MediaProjection returns black. The thumbnail Android stores for the recents stack comes out blank as well, which is the part users notice first when an app suddenly looks broken in the task switcher.

No dialog. No toast saying a capture was refused. An empty frame, and the app never learns who asked for it.

iOS solves the same problem at a different layer, and less completely. Apps poll UIScreen.isCaptured, public since iOS 11, and hide their own content when the answer is true. The system separately excludes secure text fields from capture. Signal’s iOS 27 change adds a stricter rule on top: screen recording, AirPlay mirroring and in-app capture get nothing from a chat view.

Where FLAG_SECURE stops working

None of that survives a privileged reader.

The flag is a cooperative instruction between an app and the compositor, not an encryption boundary. Nothing is encrypted. Nothing is authenticated. A window asks not to be drawn into a capture buffer, and everything above the compositor is unaffected by the request.

A capture flag describes a surface, not a secret. Accessibility services, OEM screen readers and rooted framebuffer dumps read the chat without ever asking the compositor for permission.

Accessibility is the gap that matters. A service registered for accessibility reads the view hierarchy: text nodes, bounds, roles, content descriptions. FLAG_SECURE does not touch that tree. A chat can render black in every screenshot taken of it while an accessibility client reads the whole conversation line by line, in order, with sender names attached. That is how screen readers work. It is also how a hostile assistant would work.

The camera needs no API at all. Point a second phone at the panel and the flag is a comment in a source file.

Root, or a custom ROM that patches the capture path, changes the answer again: the flag becomes advice rather than a rule. Same lesson as sandbox design, where Nitro sandboxes isolate the kernel, not your threat model. The boundary you get is the boundary the platform enforces at that layer, nothing more.

Which makes the detection matter more than the blocking. Auto-enabling screen security means Signal noticed an assistant reading the screen and reacted. I cannot read Signal’s detection code from here, so treat the mechanism as likely rather than confirmed: an accessibility-service check, or a capture-detection signal on newer Android builds. Either way the toggle flips after the reader has access. It does not revoke the permission. Revoking it is a job for the settings screen, and the app cannot do that on the user’s behalf.

Which readers each control covers

ReaderAndroid screen securityiOS 27 block
Manual screenshot or another app’s MediaProjectionBlocked, black frameBlocked
Recents thumbnailBlankedNot applicable
Accessibility service dumping text nodesNot blockedNot blocked
Camera pointed at the panelNot blockedNot blocked
Rooted framebuffer captureNot blockedNot blocked on jailbroken devices
OEM-privileged screen assistantNot reliably blockedNot reliably blocked

The bottom row is the one Signal is worried about, and it is the row the feature cannot fix. A vendor assistant with system-level screen access outranks any per-app window flag. Studies keep producing the same result from the other direction: an assistant with read access sees every app at once, which is why prompt-log research keeps finding secrets where nobody expected them.

Auditing screen readers on your own phone

The accessibility list is readable over adb without root. Two settings cover most of the exposure:

adb shell settings get secure enabled_accessibility_services
adb shell settings get secure enabled_notification_listeners

Output is a colon-separated list of ComponentInfo strings, package and class pairs. Empty output means nothing is registered. Any entry you cannot identify is a service with the ability to read the screen of every app that does not opt out, and most apps do not opt out.

Notification listeners matter for the same conversation. A listener receives message text in the notification payload, so a locked phone with previews enabled leaks more than an unlocked phone with previews off. Neither Signal’s toggle nor iOS 27’s screenshot rule touches that path.

What to change on the device instead

  • Turn off lock-screen notification previews for messaging apps, or set them to hide content.
  • Keep the accessibility list at zero entries you did not install on purpose.
  • Check which app holds the default assistant role, and whether it has screen context permission.
  • On iOS, treat the recording indicator in the status bar as a live capture warning, not decoration.
  • Assume any assistant with screen access has read access to every chat, not to one.

That last point is the uncomfortable one. People install an assistant for convenience and grant a permission wider than the feature it enables. The assistant sees the banking app, the password manager, the one-time code and the chat, and no window flag in any of them changes that. The egress problem has the same shape as Qwen3.8 trying to phone home: the software you trusted with local execution still had a path out.

One test settles it for your device

Ten minutes, one install. Register an accessibility-based screen reader, open a Signal chat with screen security enabled, and ask the reader to dump the current window. If the text comes back, the flag did nothing against that class of reader. If it comes back empty, the reader is using a capture path rather than the view tree, and the next question is what else on the phone takes the first route.

Signal is right that OS-level AI changes the threat model, and right to spend engineering time on it. What it has is the only lever an app owns: refusing to draw. Whether that holds depends on a permission screen the user controls, inside a settings menu most people open once and never revisit. Keep the toggle on. Do not mistake it for a boundary.

Providers promising deletion already gave us a preview of how this ends. Two million canary prompts found three providers keeping data they said they deleted, and not one of those promises was enforceable by the user. A screenshot flag is a stronger promise than a privacy policy, but it is the same category of promise.

Press ⌘K to search & switch themes