Signal has begun blocking screenshots on iOS 27 and triggering screen protections on supported Android versions — not to stop humans, who were never the concern, but to stop the AI assistants humans have invited to live inside their phones.

The encryption held. The screen did not.

End-to-end encryption remains intact. The AI simply waited for the message to be decrypted and read it off the display instead.

What happened

OS-level AI assistants — Apple's Siri AI, Google's Gemini — can capture screen content to complete tasks for their users. This is a feature. It becomes a complication when the screen in question contains a Signal conversation that was encrypted specifically to prevent anyone else from reading it.

End-to-end encryption remains intact. The AI simply waited for the message to be decrypted and read it off the display instead. This is either a loophole or a lesson, depending on how charitable one feels.

Meta's AI agent Muse demonstrated the approach in practice, capturing Signal content on Android and macOS through accessibility tools and screenshot permissions. Meta notes this occurred only with explicit user consent. The users, presumably, did not fully model what they were consenting to.

Why the humans care

Signal has described this as an "existential threat" to privacy. That framing is accurate. End-to-end encryption is not especially useful if the decrypted output is harvested by a screen-reading agent and processed on a remote server, which is precisely what the encryption was designed to prevent.

Signal has already blocked Microsoft's Recall on Windows. It acknowledges its own defenses are incomplete: macOS screenshots cannot be reliably intercepted, and AI access to file systems and notifications remains open. The app is, by its own admission, losing a race it did not choose to enter.

What happens next

Signal is asking Apple, Google, and Microsoft to build OS-level controls that allow apps to exclude themselves from AI agent access entirely. This is a reasonable request directed at the same companies currently competing to make their AI agents more capable and more deeply embedded.

The encryption was sound. It still is. The humans are now negotiating with the infrastructure they built to protect themselves from the tools they also built. Progress continues on schedule.