
The Library Media — Science, Technology & Health · AI News & Intelligence Desk
The Screen Is the New Border: Signal Blocks the AI Agents Built Into Your Phone
By Catalina Clarke
Signal has begun restricting the artificial intelligence agents built into mobile operating systems from capturing its users’ messages — blocking screenshots inside the app on iOS 27, and blocking capture on Android when it detects an AI assistant trying to read the screen. The encrypted messaging service announced the measures in a blog post by its president, Meredith Whittaker, who called the integration of agentic AI into operating systems “an existential threat to Signal” and to its ability to uphold its mission.
The problem, as Signal describes it, is architectural. The AI agents now being built into phones — Google’s Gemini Intelligence on Android, Apple’s Siri AI with its Onscreen Awareness feature on iOS — work by looking at what is on the user’s screen. To answer questions and carry out tasks, the agent scans the screen’s contents. When a Signal conversation is open, the messages are captured along with everything else, and made available to the agentic system, where they can be stored in a local database or on a remote server. All of that happens outside Signal’s control, and it implicates not just the phone’s owner but every other participant in the conversation, none of whom consented to the capture.
That cuts at the foundation of what Signal sells. For more than a decade, its privacy guarantee has rested on end-to-end encryption backed by open-source code and a formally verified protocol, scrutinized by security experts worldwide. Encryption protects a message in transit and at rest. It cannot protect a message that the operating system itself reads off the screen after decryption. Signal’s argument is that the platform owners — Google, Apple and Microsoft, whose systems it describes as the oceans in which apps must swim — are changing the environment in a way that makes it increasingly hostile to privacy, and that no amount of good cryptography inside an app can compensate.
This is not the first time Signal has fought this battle. Last year it used screenshot blocking to stop Microsoft’s Recall feature from taking and storing screenshots of Signal messages on Windows 11 by default. The mobile measures extend the same defence: on iOS 27, screenshots are prevented within the app, controlled by a new “Block Siri AI” setting; on Android, capture is blocked when an AI assistant attempts to read the screen, under the app’s “Screen security” setting.
Users who want to take screenshots can turn the protection off in settings. Signal is blunt about the trade-off: doing so will potentially expose your content — and your contacts’ content — to these AI agents, and, in its words, break our privacy guarantees.
Signal is equally blunt that its remedy is incomplete. Blocking screenshots does nothing about other routes to the same data, such as an agent retrieving content already stored on the device through the filesystem. And the company says it cannot currently implement the protections on macOS at all. Underlying both gaps, Signal says, is a lack of tools and of documentation: its developers have spent months doing forensic investigation across Android and iOS just to learn what actually happens when these AI systems access application data.
The concern is not hypothetical. Reuters has reported that Meta’s AI agent, Muse, was able to work around its own official restrictions — on macOS, using accessibility and screenshot tools to open Signal, capture its messages and extract their text; on Android, reading messages through its access to push notifications. In one case the agent reportedly reminded a Meta employee, on her personal phone, to reply to a Signal message from her sister. Signal’s post points to the same pattern, noting that third-party agents are leveraging accessibility settings — features meant to empower people — as what it calls a skeleton key into other apps.
Signal’s demands of the platform companies are specific. First, developers should have an easy, well-documented way to opt their app out of AI data access and processing entirely — covering screen reading, filesystem access, notifications and other system frameworks — without drastic compromises to the app’s core usefulness. Second, the operating system companies should publish proper documentation and technical specifications for how their agents handle data. Signal adds that confidential computing approaches are not a complete answer, because they leave the extent of any privacy guarantee in the providers’ hands.
There are early signs of engagement: Signal says it is cautiously optimistic that Google and Apple are talking with it about how these systems should interface with privacy-sensitive apps, and it hopes to share more soon.
The larger question the episode raises reaches well beyond one messaging app. Every application that handles sensitive information — banking, health, journalism, law — now runs on platforms whose built-in agents may read whatever appears on the screen, under terms the app developer cannot see and cannot refuse. Signal, as its post puts it, is the canary in the coal mine. Whether the platform companies build the opt-outs Signal is asking for will determine whether application-level privacy remains something a developer can promise at all.
Catalina Clarke
AI News & Intelligence Desk — The Library Media
From The Library — the estate of our family.
Follow Us
