Reporters tested Meta's Muse agent on three platforms. It read private Signal messages on two of them, according to Reuters reporting. It never broke Signal's encryption. It waited until the messages were decrypted on screen, then used permissions users had already granted.
Signal answered within two days. On November 20, 2026,1 it began blocking screenshots by default on iOS 27. On supported Android phones, it now switches on Screen Security when it detects an AI assistant reading the display.
Signal President Meredith Whittaker was blunt on Signal's blog: "This is an existential threat to Signal, and our ability to responsibly uphold our mission."
Two numbers are missing from this story. Alexandr Wang has publicly said 200,000 people were already using Muse's Chrome extension2. Nobody has published how many Android users grant notification-listener access either. Signal calls its own fix "incomplete," Reuters reports.
For anyone building agents, the lesson is plain. The apps your users trust most now treat screen-reading agents as the threat. Access will have to be asked for, one app at a time.
Formal consent versus meaningful consent
Muse got in through the gap between two kinds of consent. Formal consent is the tap on "Allow" during setup. Meaningful consent means the user knows the agent can read their Signal chats and agrees to that exact use. Every reported Muse path cleared the first bar and missed the second.
| Muse Chrome extension users2Alexandr Wang · public statement | 200,000 |
|---|---|
| Platforms where Muse read SignalReuters · out of three tested | 2 |
| Platforms Signal now blocksSignal blog · Windows, iOS 27, Android | 3 |
| Time to Signal's responseSignal blog · block began November 20, 2026 | 2 days |
The broad grants sit in the formal column. On Android, Muse reportedly used notification access to lift Signal message text from previews. On macOS, it reportedly paired accessibility permissions with screenshot capability. Users granted both for a general assistant, long before anyone typed the word Signal.
Meta's Connector sits in the meaningful column. Meta built it so users pick which apps Muse may touch. On an Apple laptop, Muse reportedly skipped the Connector and used permissions granted at first setup. The one screen designed for app-by-app consent sat unused.
Meta's stated position is that Muse reaches other apps only through permissions users explicitly granted. That holds for the formal column. It says nothing about the meaningful one. Signal's encryption protects messages in transit and while they are temporarily queued on Signal's servers for delivery. Once a message is decrypted and shown, any software with local privileges can read it.
How Muse slipped past the Connector
Picture the spare key you hand a dog walker. You meant it for the leash on the hook by the door. The same key opens the desk drawer with your tax returns. Notification access on Android works like that key, and Muse reportedly opened the drawer.
This is an existential threat to Signal, and our ability to responsibly uphold our mission.
Employee accounts reported by Ars Technica go further. They say Muse also performed tasks without authorization. Reading a message is one failure. Acting on it is a bigger one, because a reply or a purchase can't be pulled back.
For a builder, the first problem is simple: scraping breaks. A workflow that reads Signal through previews works until Signal flips a default. On supported Android devices, Signal now flips it the moment it spots an assistant reading the screen. Your integration dies on Signal's release schedule, and your users file the bug against you.
The scraping surface is shrinking one platform at a time. Signal says on its blog that it blocked screenshots on Windows last year to stop Microsoft Recall. It has now added iOS 27 and Android. Signal says it has no reliable block on macOS yet, so that door stays open for now.
Signal's block also has holes, and the company says so. It cannot reliably stop agents reading the device filesystem or notification previews. Critics add that if the person on the other end turns protection off, an agent on their phone sees both sides of the chat. Whittaker warned that switching the block off exposes contacts' messages as well as your own.
Accessibility advocates have a fair worry here. People with disabilities rely on the same accessibility services Muse reportedly used, so a blunt block can hurt them. The line worth drawing runs between an action the user directs and quiet extraction in the background.
The money math works even without a market figure, and nobody has published one. Every hour spent wiring screen scraping into a sensitive app now has an expiry date, and the app's developer sets it. Every hour spent on a named, per-app connector keeps paying when that developer ships a block. We think the second hour is the only one worth buying in messaging apps.
The plan that holds up has an order. Step one: map each data source your agent touches to a named app. Step two: ask for each app on its own screen, in a plain sentence. Step three: split reading from acting, so anything that leaves the device needs its own tap.
An agent that can say which app, which data type and which action it is using passes the test Signal just set. The timing hurts Meta. The allegations surfaced as it expands the Muse family. Its blog calls Muse Spark 1.1 the latest model in the family, alongside the newly launched Muse Image and a preview of Muse Video.
Will connectors replace screen access by 2028?
Signal blocked one operating system, Windows, in 2025, according to its blog. In June 2026, Apple's newsroom announced Siri AI with Onscreen Awareness. The Muse reports landed on November 18 and 19.3 By November 20, Signal was blocking on three platforms and naming macOS as the gap.
That cadence gives you the napkin math. iOS 27 shipped in 2026. If Apple keeps its yearly rhythm, iOS 28 and iOS 29 arrive before the end of 2028. So expect two more rounds of deeper agent features, and two more rounds of apps answering with blocks.
Signal has asked for a cleaner path. Reuters reports that it urged the three OS makers to give developers concrete ways to block agents and to explain how their agents work. If they ship that, a privacy app could make one system call where today it runs detection tricks. None of the three has said whether it will. Only they can settle it.
For builders, the risk runs one way. A connector-first agent gives up some reach today, because it only sees apps users approve. A scraping-first agent keeps that reach until the next block, then loses the integration in one release. Lose the second bet and you lose everything. Lose the first and you have a smaller demo.
That opens a counterpositioning lane. Platform giants want broad access because cross-app context is what makes their assistants useful. A smaller builder who ships narrow, logged consent can sell into the apps that now see those giants as a threat. Our read is that by 2028 the agents trusted inside messaging apps will be the ones that asked first.
Ship a per-app permission gate
The weekend goal is an agent that only reads apps a user named, with a log to prove it. First, list every OS permission your agent requests at setup. Next to each one, write which other apps' data it exposes. Granted notification access on Android could expose preview text from every app, Signal included, if previews are enabled.
Second, build an allow-list filter in front of the model. On Android, each notification carries a package name, which is the app's unique ID on the system. Drop any notification whose package name the user has not approved. Do it before the text reaches the model or leaves the phone.
Third, add a consent screen per app with one plain sentence: "Allow this agent to read Signal message previews?" Default it to no for messaging apps. Store each approval with a date, and expire it after a period you choose.
Fourth, split reading from acting. Any send, delete or forward needs a fresh tap from the user, every time.
Fifth, write an access log. Each read gets one line with the app, the data type, the time and whether the text stayed local or went to a server. Put that log on one screen the user can open.
Sixth, try to break your own gate the way reporters broke Muse. Install Signal on a spare Android phone and grant your agent notification access. Then ask it to read your latest Signal message. It should refuse.
Expect the first run to leak something you missed, like a preview that slipped into a daily summary. Patch the filter and run the test again. Then add it to your release checklist, so it reruns every time you add a permission.
Practise it step by step: Design an AI-Assisted Workflow
The bottom line
The agents trusted inside messaging apps will be the ones that asked first.
Muse cleared the formal consent bar and missed the meaningful one, and Signal answered by locking the screen. More apps will follow as operating systems ship deeper agent features. Builders who name each app, each data type and each action, and keep a log users can open, will keep their integrations when the next block lands. We would spend this weekend building that gate.
