All News
signalprivacyai-agentsapplegoogleandroid

Signal blocks screenshots to shut OS-level AI agents out of your chats

Signal turns on screenshot blocking in iOS 27 and Android to keep Gemini and Siri AI away from your chats, and asks Apple, Google and Microsoft for opt-outs.

Vlad MakarovVlad Makarovreviewed and published
7 min read
Signal blocks screenshots to shut OS-level AI agents out of your chats

Meredith Whittaker's post does not open like a product announcement. Published on 20 November, it starts with the mission and arrives, within three sentences, at an operational decision: Signal is restricting the AI agents built into mobile operating systems from capturing and storing your messages. The mechanism is screenshot blocking, on by default in Signal for iOS 27 and switched on automatically in the Android app when it notices an assistant reaching for the screen. Whittaker's warning to anyone who disables it: doing so "will potentially expose your (and your contacts') Signal content to these AI agents, and break our privacy guarantees."

That is a strange position for a privacy tool. Signal has spent a decade arguing its guarantees rest on auditable code and a formally verified protocol. Now the last line of defence is a switch that removes a normal phone feature, and the switch has holes.

What the app now does, and to whom

The two platforms are handled differently, and the footnotes explain why. On iOS 27 the protection is unconditional: screenshots inside Signal stop working, with a "Block Siri AI" toggle carrying the restriction in Settings. On Android it is reactive. Signal writes that many Android builds exist beyond Google's Pixel images, and that on many of them a screenshot attempt cannot be detected once screenshots are already disabled. Rather than ship a version where screenshots quietly break, it enables the "Screen security" setting itself when an assistant reaches for screen data.

The agents being fenced off are named. Google is folding Gemini Intelligence, backed by Google Gemini, deep into Android; Apple has built Onscreen Awareness into Siri AI. Neither is an app a user opens. Both sit underneath the app, which is the problem, and Whittaker reaches for a metaphor: the major operating systems are "the oceans in which Signal, and other applications, must swim," and they "dictate the conditions of life for the application ecosystem." Her conclusion comes without a hedge: "This is an existential threat to Signal, and our ability to responsibly uphold our mission."

Recall was the rehearsal

Signal has fought this battle before. Last year it used screenshot blocking against Microsoft's Recall on Windows 11, which captured the screen every few seconds by default. A footnote spells out what that meant here: optical character recognition turns the frames into text, the text lands in a searchable on-device database, and the database can hold your Signal messages indefinitely, including the disappearing ones the app promises to delete.

The Windows fight had a clean shape: a feature, a stated default, a patch. The mobile version is messier, because collection happens inside the operating system rather than next to it. Once an agent has lifted a conversation off the screen, Signal has no further say: the data goes to "whatever local database or remote server an OS-level AI agent is using at that moment." The loss is not confined to the person holding the phone; every participant inherits it.

Where the patch stops

Signal is candid about the limits, worth noting in an industry where security claims arrive without footnotes. Blocking screenshots does not close "other more sneaky ways AI agents could potentially access application data," starting with content already on the device in the filesystem. An agent with file access does not need to photograph anything.

Then there is macOS, where Signal cannot do this at all. The APIs that would allow screenshot protection were deprecated in favour of protections that only work for streaming media and other DRM content. Continuous screen capture can be restricted; single screenshots, the route most agents use, cannot. Signal's only advice for macOS 27 is to add Signal to the Excluded Apps list in System Settings and then Siri.

The company also spent months on forensic work across Android and iOS, and says plainly that the documentation was not there to read. "Without better documentation and clear technical specifications," it warns, "it will be increasingly difficult for Signal to provide robust privacy and security to the many millions of people who rely on us." That sentence cuts both ways.

Two asks, and who has to say yes

The demands go to Apple, Google and Microsoft, and are narrow enough to be plausible. First, an easy, documented way to opt an app out of AI data access and processing, covering screen reading, filesystem access, notifications, APIs and other OS frameworks, without gutting the app's core utility. Second, real specifications for how these agents function, especially around data access.

The post pre-empts the obvious counterargument: confidential computing is not a fix, because those approaches "still leave the power to determine the extent of privacy guarantees (if any) in the hands of these providers." It extends the worry to third-party agents such as Meta's Muse, which lean on permissions users are pushed to grant at setup, and notes that accessibility settings meant to help people use their phones are being used "as a skeleton key" for agents. Signal is not alone in pushing: the EFF's October win let users exclude apps from Siri AI, a small version of the first ask. Coverage from Reuters and The Verge framed this as Signal versus the assistants, which understates the oddity: the company is asking its competitors' landlords to write better tenancy rules.

The part nobody outside can measure

The evidence that OS agents scan the Signal screen comes from Signal itself. On Android the app detects an assistant attempting to read the screen; that detection is the trigger, and it is also the entire public record of the behaviour. Neither Apple nor Google has published a specification describing what Onscreen Awareness or Gemini Intelligence captures, how long it keeps it, or where it sends it. An outside researcher cannot measure the scanning, because the systems are closed and undocumented.

What shipped this week is therefore a precaution, not a proven breach. On iOS the blocking does not depend on detecting anything: screenshots are off by default whether or not an agent ever looks at the screen, and the user pays in lost functionality. That is a defensible trade for a messenger whose threat model includes a hostile platform, but it is a trade taken on a risk Signal cannot quantify in public. The claim that agents are reading your messages and the claim that they could read them produce the same patch and very different journalism.

The Hacker News thread, at 569 points and 333 comments, split along a familiar line: people who consider app-level resistance futile against an OS vendor, and people who consider doing nothing worse. Both camps are arguing about the same missing artifact, an independent account of what the agents collect.

What would settle it

Two things would move this from a well-argued position to a verifiable one. The first is the remedies Signal asks for: an opt-out that is documented, testable and uniform would let any auditor confirm what an agent can and cannot reach. The second is an independent audit of the capture pipeline, the kind of work that turned Recall from a marketing feature into a documented liability. Until one exists, every claim here, including Signal's, rests on the party making it.

For users the practical steps are unglamorous. Keep the protection on if you use Signal for anything sensitive, and accept that you will not be screenshotting conversations. On a Mac, add Signal to the Excluded Apps list under System Settings and Siri. On a phone, remember the toggle exists, and that turning it off is a decision about other people's messages too.

The forecast is that this fight keeps recurring, in different apps, with the same asymmetry: the platform owner sets the default, the developer patches around it, and the user trusts whoever seems more credible. Whittaker calls Signal "the canary in the coal mine." A canary is useful, but it is also a warning that the air has already changed.

Related Articles

Scroll down

to load the next article