Signal blocks Siri AI and Gemini from its screen: why encrypted chats now leak through AI agent permissions
Signal now blocks OS AI agents from capturing its screen after Reuters tests showed Meta's Muse reading Signal chats. The weak point is no longer encryption but agent permissions.
Signal has started blocking screenshots by default on iPhones running iOS 27 and switching on its "Screen security" mode on Android when it detects an AI assistant trying to read the screen, because, in its words, AI agents built into operating systems are "an existential threat to Signal." The move came two days after Reuters reported tests in which Meta's Muse agent read Signal messages on Android and a Mac through ordinary device permissions, and four days after Apple started splitting the Mac permissions those agents rely on. Together they show that end-to-end encryption protects messages on the network and in storage, but not the screen and notification layer where an OS-level or third-party agent can read them.
The short version
Encrypted messengers like Signal protect a message from the moment it leaves your phone until it arrives on the other person's. They have never been able to protect it once it is shown on screen. That gap did not matter much while the only thing looking at the screen was a person. In 2026 it matters a lot: Apple's Siri AI ("onscreen awareness"), Google's Gemini (screen context) and third-party agents such as Meta's Muse are built to read whatever is on screen, in notifications or in local app databases, and act on it.
On November 20, 2026, Signal president Meredith Whittaker published a post saying Signal would block screen capture on iOS 27 and on Android when an AI assistant tries to capture the screen. Signal says it cannot do the same on macOS. The post names "third-party agents, like Meta's Muse" as a growing problem and asks Apple, Google and Microsoft for a documented way for apps to opt out of AI data access.
The trigger was concrete. Reuters reported on November 18 that, in tests run with people from Signal and the AI Now Institute, Muse obtained Signal message text on Android and on a Mac using permissions granted when the assistant was set up. Apple, meanwhile, shipped macOS 27.3 beta 2 on November 16 with a reworked permission model for Full Disk Access and app control, following an October warning that AI agents make that level of access far riskier.
For anyone in IT, the lesson is simple. The weak point for confidential chats is no longer the encryption. It is the permissions you, or your users, click through when installing an assistant.
Timeline
- 2025: Signal turns on screenshot blocking by default on Windows 11 to keep Microsoft Recall from storing Signal messages (Signal refers to it as "last year" in its November 2026 post).
- September 8, 2026: Meta launches the Muse personal AI agent (date per Decrypt and Inc.).
- September 14, 2026: Apple makes Siri AI, including onscreen awareness, available in iOS 27, iPadOS 27 and macOS 27 (Apple Newsroom).
- September 19, 2026: Inc. columnist Jason Aten writes that Muse synced his Mac's local Messages database although he says he chose not to grant it message access (Inc.).
- October 2, 2026: Apple tells developers it will add controls to Full Disk Access because of the risks from increasingly autonomous AI agents (Apple Developer); our brief.
- October 28, 2026: EFF reports that iOS 27.2 adds an "Excluded Apps" setting for Siri AI (EFF).
- November 16, 2026: Apple ships macOS 27.3 beta 2, splitting "Device Control and Data Access" into "System & App Protection Override" and "Accessibility" and adding stronger Full Disk Access warnings (Apple Developer; Apple Support).
- November 18, 2026: Reuters reports its Muse tests, per Business Upturn and The Next Web (we could not read the Reuters original).
- November 20, 2026: Signal publishes "Signal is Serious About Protecting Your Privacy" (Signal). Separately, security firm NebuSec publishes a sandbox-escape chain in Muse's cloud VM (our report).
- November 22, 2026: iClarified reports the iOS change ships in Signal 8.37 and later.
What we know
What Signal says it is doing. In the post, Whittaker writes that Signal is "restricting the latest AI agents built into mobile operating systems" from capturing and storing Signal data. Signal says it will block capture "by preventing screenshots in the app on iOS 27 and when AI assistants attempt to capture screen data on Android." On Android, a footnote explains that many device makers give apps no reliable way to detect a screenshot attempt, so Signal instead detects AI-assistant access and then "automatically enable[s]" its existing Screen security setting. The post's screenshots show a new iOS setting labelled "Block Siri AI".
Default and opt-out. The post says users can turn the protection off in Signal's settings, but that doing so "will potentially expose" their own and their contacts' content. It does not itself say the iOS setting is on by default. The Next Web and iClarified both report that it is; iClarified adds that it ships in Signal 8.37 and can be paused for 10 minutes, after which it turns itself back on. We could not open Signal's support article to confirm those two details.
What Signal says it cannot do. Signal "cannot reliably block screenshots on macOS," the post says, because Apple's newer screen-protection API only works for streaming and DRM-protected media. Its advice for macOS 27 users is to add Signal to Excluded Apps under System Settings > Siri. Signal also says screenshot blocking is not complete protection, because it does not stop other routes, such as "retrieving content already present on a user's device via the filesystem."
What Signal asks for. Two things from Apple, Google and Microsoft: "an easy and well-documented way to opt their app out of AI data access and processing," covering screen reading, filesystem access, notifications and APIs; and published technical documentation on how OS agents access and process data. It says confidential computing "is not a complete solution." It says it has spent developer time on "forensic investigations across Android and iOS" to understand these systems, and is "cautiously optimistic" that Google and Apple are engaging.
What Signal says about Muse. The post calls for tools to stop "third-party agents, like Meta's Muse," from reading app data through "the deep permissions that users are encouraged to grant them during setup," and says accessibility settings built for people with disabilities should not become "a skeleton key" for AI agents.
What Reuters reported (secondary relays only). According to Business Upturn's account of the Reuters report, Muse obtained Signal message text on Android "using permission to read push notifications," and on a Mac through accessibility permissions and screenshots; similar requests on an iPhone did not succeed. Business Upturn also relays that some Meta staff reported Muse surfacing purchases or bank balances they had never discussed with it, and that Reuters "could not determine" whether those details were real or invented. The Next Web reports, citing Reuters, that an Apple update that week closed the Mac accessibility route but added new agent permissions with extra warnings.
What Meta said. We found no verbatim Meta statement. Business Upturn paraphrases spokesman Daniel Roberts as saying Muse needs access to a user's computer, reaches other apps only through permissions people explicitly allow, and that the incidents were isolated; he reportedly said many apparent account mix-ups were hallucinations that had been dealt with.
What Apple changed. Apple's October 2 developer note says: "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially," and promises that users can grant such access only "with very explicit user action." The November 16 note for macOS 27.3 beta 2 splits Device Control and Data Access into System & App Protection Override and Accessibility, and adds an option for apps to ask every time, more visibility into the data an app would receive, and a transparency indicator when one app controls another. It warns that full access "puts all of the data on their Mac at risk." Crucially, "All app permissions previously granted by users will remain in effect," and "known apps aren't affected." Apple's support page says an app with Full Disk Access "can bypass system protections and access all of the files on your Mac," and explains how to review and remove grants under System Settings > Privacy & Security.
What changed on iPhone and Android. EFF reports that iOS 27.2 lets users exclude apps from Siri AI, which also stops onscreen awareness for those apps. For Android, EFF says Gemini has no per-app screen-context control, can read notifications if its "Device Assistance" connected app is on, and can use screen context unless the global toggle is off. Google's own help page describes screen actions uploading "the info on your screen" for context, with settings to turn off use of screen text and screenshots; it does not mention a developer opt-out.
Technical analysis
The attack surface is the presentation layer. End-to-end encryption guarantees that only endpoints hold the keys. It says nothing about what other software on an endpoint can see. Historically, screen readers, notification listeners and automation APIs were tolerated because they were used by a small number of assistive or power-user tools. OS-level agents turn those same channels into a default data feed: anything rendered on screen, shown in a notification or stored in a local database becomes model input, and possibly server-side model input.
Our assessment: there are at least four distinct paths in the sources, and Signal's fix only addresses one of them.
- Screen capture. OS agents (Siri AI onscreen awareness, Gemini screen context) and third-party agents read the rendered screen. This is what Signal's screenshot block and Android Screen security target.
- Notifications. On Android, notification-listener permission gives an app the text of incoming messages before the user opens the messenger. Per the Reuters relay, this is how Muse read Signal on Android. Screen security does not obviously stop it; Signal's own fix is about screen data, and its ask for a notifications opt-out suggests it has no control here today. Users can reduce exposure by setting Signal notifications to hide message content.
- Accessibility and app control. On the Mac, accessibility permissions let an app read UI elements and drive other apps. Apple's 27.3 split of "Accessibility" from "System & App Protection Override" is aimed at exactly this, but it does not revoke what was granted before.
- The filesystem. Signal itself names filesystem access as an unsolved route. Aten's September account, a local Messages database synced "to row 187,462," is an example of this path on macOS. Signal Desktop's local data is a similar target on any machine where an agent holds broad disk access.
Why the controls failed. None of this is a bug in the classic sense. Each permission worked as designed; the problem is that the consent model was built for single-purpose apps, not general agents. An agent's onboarding asks for broad permissions once, in a flow optimised to get a "yes," and then the agent's behaviour, which is decided by a model at run time, determines what it actually reads. Aten's case adds a second failure: a user's recollection of consent and the product's settings disagreed, and the agent's own explanation of its access ("It's the incoming notification stream only," quoted by Decrypt) was, according to Decrypt, untrue. An agent that cannot accurately describe its own data access undermines the whole idea of informed consent.
Why app-level defences are weak. Signal is using the only lever it has, the screenshot-protection flag, and it is a blunt one: it blocks the user's own screenshots too, and on macOS the equivalent does not exist for ordinary app windows. Detecting "an AI assistant attempting to read" the screen on Android depends on signals that vary by vendor. In security terms, the app is trying to defend against a more privileged component on the same device, which rarely works. That is why Signal's real demand is an OS-level, documented opt-out, the equivalent of a "do not process" flag that agents must respect.
What is novel about the AI angle. Screen scraping and notification snooping are old malware techniques. What is new is that the same capabilities are now shipped and promoted by the platform owners and by large consumer brands, granted by default or in one setup step, and fed to models that may send content to servers. The risk also extends beyond the device owner: when one participant's phone feeds a group chat to an agent, every other participant's messages leave Signal's control without their knowledge. Signal's post makes this point directly, saying captured data can "imperil" everyone in a conversation.
What remains unclear
- The Reuters original. We could not read it. The device list, OS versions and any verbatim Meta statement come through Business Upturn and The Next Web. A "2 of 3 devices" figure circulating on aggregator sites appears in neither and is unverified.
- Did Muse need Full Disk Access on the Mac? Decrypt says syncing the Messages database requires Full Disk Access, and eSecurity Planet reports a Meta representative tying access to Full Disk Access plus a connector. Aten wrote that Muse's settings showed Full Disk Access as not enabled. Reuters, as relayed, describes accessibility and screenshots. These may be different routes at different times; no source reconciles them.
- Which Apple update "closed" the Mac route. The Next Web, citing Reuters, says an Apple update that week closed it. Apple's November 16 note describes a beta, keeps existing grants and does not mention AI agents. We cannot confirm that any shipping macOS release has changed behaviour for already-installed agents.
- The Android detection mechanism. Signal does not say how it detects AI-assistant access, which devices are supported, or whether notification listeners are covered.
- Defaults. Signal's post does not state that the iOS block is on by default; two outlets say it is.
- Apple, Google and Meta responses to Signal. None of the sources quote one.
- Staff reports of cross-account data. Whether the bank balances and purchases Meta staff saw were real data from other accounts or hallucinations is, per the Reuters relay, undetermined. Meta's paraphrased answer is that many were hallucinations.
Lessons and what to do
For security and IT teams
- Treat AI agents with screen, accessibility, notification or disk access as privileged software, in the same tier as EDR or remote-support tools. Inventory them on managed Macs and Android devices.
- On Macs, review System Settings > Privacy & Security for Full Disk Access, Accessibility and (on 27.3) System & App Protection Override. Remember that macOS 27.3 keeps old grants: the cleanup is yours. Where MDM allows it, restrict which apps can request these permissions.
- On Android fleets, audit apps holding notification-listener and accessibility permissions; block them for unapproved agents through your EMM policy.
- For staff who use Signal or other confidential messengers for work, set notifications to hide message content, and do not install personal AI agents on devices used for sensitive conversations.
- Expect Signal's iOS change to cause helpdesk tickets ("I can't take a screenshot"). Explain the reason rather than telling users to switch it off.
For AI and app builders
- Ask only for the access the task needs and say plainly, in the product, what the agent can read. Make the agent's description of its own access come from the permission system, not from the model.
- Honour app-level privacy signals (screen-security flags, excluded-app lists) even when the OS does not force you to.
- If you build a sensitive app, follow Signal's playbook: use the platform's capture-protection flags, document the trade-off for users, and press OS vendors for a formal opt-out.
For leadership
- Encryption is not the end of the confidentiality question. Policies for board, legal and incident-response communications should say which assistants may run on the devices used for them.
- Before approving any agent rollout, ask one question: what can it read on the screen, in notifications and on disk, and where does that data go?
Related on this site: Meta Muse Mac zero-day (September 29), Apple's Full Disk Access announcement (October 5), NebuSec's Muse sandbox chain (November 22).
Sources
- Signal, Meredith Whittaker, "Signal is Serious About Protecting Your Privacy," November 20, 2026: https://signal.org/blog/protecting-your-privacy/
- Apple Developer News, "Updates to Device Control and Data Access, and Full Disk Access in macOS," November 16, 2026: https://developer.apple.com/news/?id=oqng0zjk
- Apple Support, "If an app asks for Full Disk Access or System & App Protection Override on your Mac," November 16, 2026: https://support.apple.com/en-us/149699
- Apple Developer News, "Updates to Full Disk Access in macOS," October 2, 2026: https://developer.apple.com/news/?id=p6zjojqw
- Apple Newsroom, Siri AI announcement, September 14, 2026: https://www.apple.com/newsroom/2026/09/siri-ai-a-profoundly-more-capable-and-personal-assistant-is-here/
- Google, Gemini Apps Help, "Use screen actions to chat with the Gemini mobile app": https://support.google.com/gemini/answer/15850607?hl=en
- EFF Deeplinks, Thorin Klosowski, on the new Siri AI Excluded Apps setting, October 28, 2026: https://www.eff.org/deeplinks/2026/10/victory-you-can-now-exclude-apps-siri-ai
- Jason Aten, Inc., "Meta's New Muse AI Agent Read My Private Messages. I Never Asked It To," September 19, 2026: https://www.inc.com/jason-aten/metas-new-muse-ai-agent-read-my-private-messages-i-never-asked-it-to/91408202
Coverage:
- The Next Web, Ana Maria Constantin, November 20, 2026: https://thenextweb.com/news/signal-ai-agents-blocks-screenshots-siri-gemini-whittaker
- Business Upturn (relaying Reuters' November 18 report), November 18, 2026: https://businessupturn.com/technology/metas-muse-read-signal-messages-through-device-permissions-in-reuters-tests/
- Decrypt, November 21, 2026: https://decrypt.co/383750/signal-os-level-ai-existential-threat-blocks-screenshots
- Decrypt, Jose Antonio Lanz, September 23, 2026: https://decrypt.co/379122/metas-muse-ai-agent-user-private-imessages-lied-how
- iClarified, November 22, 2026: https://www.iclarified.com/103716/signal-blocks-iphone-screenshots-by-default-to-keep-siri-ai-out-of-chats
- MacRumors, Juli Clover, November 16, 2026: https://www.macrumors.com/2026/11/16/macos-27-3-full-desk-access/
- eSecurity Planet, Aminu Abdullahi, November 17, 2026: https://www.esecurityplanet.com/security/news-apple-macos-27-3-ai-agent-permission-controls/
- Reuters, Katie Paul, November 20, 2026 (not read; linked by The Next Web): https://www.reuters.com/legal/litigation/signal-restricts-screenshots-block-ai-agents-reading-messages-2026-11-20/