Encrypted messaging app Signal is moving to limit screenshots in response to a growing concern: operating-system AI agents that can inspect what appears on a user’s screen.

Signal President Meredith Whittaker said the company will prevent screenshotting in its iOS 27 app and restrict captures when AI assistants try to collect screen data on Android. The change responds to a class of features designed to use on-screen content as context for completing a person’s request.

Why Signal sees a new privacy gap

Signal’s encryption protects messages as they travel between participants and helps ensure that conversations are readable only by their intended recipients. But protection inside an app does not necessarily extend to what happens after text has been displayed on a device screen.

That distinction is central to Signal’s decision. Whittaker argued that once sensitive content is captured by an operating-system-level AI feature, it can become available to the local database or remote server used by that feature. Signal does not control that capture process, even if the conversation itself remains encrypted within the app.

The concern is not limited to a single assistant. Google and Apple are developing agentic features, including Gemini Intelligence and Siri AI’s Onscreen Awareness, that use screen information as context. Such capabilities can be useful for carrying out tasks, but Signal says they create a privacy challenge for third-party communication services whose security assumptions depend on messages staying inside a protected environment.

A repeat of the Recall response

This is not Signal’s first attempt to address screenshots generated outside its app. The company previously blocked screenshots in response to Microsoft Recall, the Windows 11 feature that could capture images of activity on a device, including Signal chats.

The latest restrictions broaden that response from a concern about a particular desktop feature to a broader concern about AI systems embedded more deeply in operating systems. In Signal’s view, the practical issue is straightforward: highly secure encryption provides limited protection if another layer of software can copy the readable conversation from the display.

The decision also illustrates how AI product design can change the operating conditions for other software. Earlier generations of consumer AI were often accessed as separate chatbots or services. Screen-aware assistants place more capability inside the operating system and potentially closer to data shown by every app on the device.

Security promises meet screen-level access

Signal’s move comes as the service continues to emphasize the strength of its cryptographic protections. The source report noted that researchers at Google DeepMind recently used AI agents to formally verify Signal’s code, building on Signal’s own work to demonstrate properties of its encryption.

That kind of verification addresses whether the system protects information according to its intended design. Screen capture is a different problem. If messages can be read by an assistant through the screen, the question is no longer whether the encrypted channel was broken. It is whether private content was copied after decryption for the recipient.

This distinction matters for users because security features are often discussed as if they apply uniformly across a device. In practice, protection depends on several layers: the service, the operating system, the permissions granted to other software, and the ways content can be exported or recorded. Signal’s restrictions are an attempt to reduce one of those paths.

What the change signals for AI assistants

Signal’s decision does not settle the larger debate over screen-aware AI. Operating-system assistants are being built to understand what users are doing and to provide help without requiring every detail to be manually copied into a prompt. But that convenience raises questions about how much sensitive context is collected, where it is stored, and which parties can access it.

For secure messaging services, those questions have immediate consequences. A conversation can include personal, professional, financial, or safety-sensitive information involving more than one person. Signal argues that participants should not lose control over that information simply because a recipient’s device is running a screen-reading assistant.

The restrictions are therefore a defensive product choice as well as a statement about responsibility. Signal is trying to preserve its privacy model in an environment where AI features can observe content beyond the boundaries of individual apps. Other privacy-focused services may face similar decisions as operating systems add more agentic capabilities.

For now, Signal’s position is clear: encryption remains essential, but it is not sufficient if the operating system can turn a private conversation into AI-readable screen data. The company’s screenshot limits are meant to close that gap before screen awareness becomes a routine part of everyday computing.

This article is based on reporting by ZDNET. Read the original article.

Originally published on zdnet.com