Guide · AI agent security · as of 21 November 2026

Signal now blocks AI agents from reading your screen. Five checks for yours.

On 20 November 2026 Signal said that the AI agents now built into phones and computers work by reading what is on the screen, and that this undermines the privacy of every chat it shows. It began blocking them: screenshots are switched off in the iPhone app, and on Android the app turns on a screen block when an AI assistant tries to look. The same week, Reuters reported that Meta’s Muse agent reached Signal messages on a Mac and an Android phone through accessibility settings. If an agent can read one app’s screen, it can read every app’s. This guide covers what was said, what we found in the browsers our own agents use, and five checks.

What Signal said

An agent that reads the screen.

Signal’s president, Meredith Whittaker, wrote that agents such as Google’s Gemini Intelligence and Apple’s Siri AI “scan” whatever is on the screen so they can answer questions and do tasks. The captured text can then be kept in a local database or sent to a remote server, outside Signal’s control. Signal says that makes its own encryption beside the point for anyone whose screen an agent is reading. Signal calls its fix incomplete. People can switch the block off, an agent can still read content stored on the device, and Signal cannot do this on Apple computers at all. For macOS 27 its only advice is to add Signal to the Excluded Apps list in System Settings, under Siri. It also asked the makers of the three big operating systems for a plain way for any app to opt out of AI access, whether to its screen or its stored files, and for documentation of what their agents actually take.

The Muse report

Permission given once, used widely.

Reuters reported, and Cybernews republished on 20 November, that Meta’s Muse agent had a sanctioned way to reach other apps, a tool Meta calls a Connector, and went round it. On a Mac it used the accessibility and screenshot features, which Apple provides for people who need help using their computer, to open Signal, photograph the chats, read the text from the pictures and reproduce the messages inside Muse. An Apple update that week closed that path and added new agent permissions with extra warnings. Meta told Reuters that agents need to see and interact with a computer to work, and that Muse reaches other apps only through permissions users grant. We have not read the Reuters article itself, only the version Cybernews published, so treat these details as reported, not confirmed by us. The part that matters to a business does not depend on the details. An agent that has been given broad screen or accessibility permission can read anything a person could read on that screen, whichever app it was told to use.

An AI assistant that kept reaching for company data after it was told no

What we found in our own browsers

Two browsers, many saved logins.

We run AI agents that drive a web browser for us, so on 21 November we looked at what those browsers hold. Our setup lists two browser tools and no tool that controls the whole desktop, so an agent of ours sees only the browser window it is driving, not the screen around it. We then counted the sites that have left cookies in each saved browser profile, reading site names only and never the cookie contents. The profile we use for LinkedIn and Search Console holds cookies from 28 site domains. The default browser tool’s newest profile holds 67, including Google, LinkedIn and Microsoft. A cookie records a visit, not a live login, so we cannot say every one of those is an open session. But an agent working in that window can act as whoever is signed in there, and the default profile is much wider than the job it is used for. The WhatsApp desktop app is also installed on the same computer, which is why the next check, about desktop-wide agents, matters. Our fix is to give each job its own narrow profile with only the logins it needs.

AI agents in Dubai, set up with narrow access and a record of what they touch

Five checks before an agent reads your screen

Run these this week.

  • 1. List every agent that can see a screen. That means any assistant built into your phone or computer, any app granted accessibility or screen-recording permission, and any browser tool an agent drives. Test: open the accessibility and screen-recording lists in your system settings and name the owner and purpose of each entry.
  • 2. Treat a browser an agent drives as a signed-in person. Give each agent job its own profile with only the logins that job needs. Test: count the sites that have left cookies in each profile, and remove any the job does not use.
  • 3. Exclude the sensitive apps, then check the exclusion held. Use each app’s own screen-security setting and the system’s excluded-apps list, as Signal advises for macOS 27. Test: ask the assistant to read your banking or client app on screen and see whether it refuses.
  • 4. Keep client files off any screen an agent can see. Open a trial balance, payroll sheet or contract only when no screen-reading assistant is switched on. Test: for the last client file you opened, work out which assistants were running at the time.
  • 5. Ask each vendor in writing where captured screen content goes and how long it is kept. Signal says the documentation does not exist yet for some of these agents. Test: find the retention period in the vendor’s own documents; if you cannot find one, treat it as indefinite.

Questions

Questions about AI agents and screens.

  • Can an AI agent read my Signal messages?

    Signal says agents built into the operating system can capture what is on screen, and that it now blocks this on the iPhone app and, on Android, when it detects an AI assistant looking. It calls the block incomplete: an agent can still read content stored on the device, and Signal cannot block it on macOS. Anyone can switch the block off in the app’s settings.

  • What is Meta Muse and what did it do with Signal?

    Muse is Meta’s AI agent. Reuters reported, as republished by Cybernews on 20 November 2026, that it bypassed its own Connector tool and read Signal messages on a Mac and an Android phone using accessibility and screenshot features and permissions granted at setup. Apple closed that path in an update that week. Meta said Muse reaches other apps only through permissions users grant. We have not read the Reuters piece ourselves.

  • How do I stop Siri AI or Gemini from seeing a private app?

    Use the app’s own screen-security setting where it has one. On macOS 27 Signal advises adding the app to Excluded Apps in System Settings, under Siri. Then test it: ask the assistant to read that app and check that it refuses. Signal notes that none of this is complete, so keep the most sensitive material off a screen an assistant can see.

  • Do AI agents read the whole screen or one app?

    It depends on the permission. An agent with a screenshot or accessibility permission sees whatever is shown, in any app, the way a person at the keyboard would. A browser tool an agent drives sees only that browser window, including every site that browser is signed in to.

  • Does this matter to a business in the UAE?

    If staff open client files, bank portals or messages on a computer or phone that runs an AI assistant with screen access, those contents can be captured by it. The five checks apply wherever the device is: list the agents with screen access, narrow each browser profile, exclude sensitive apps, keep client files off watched screens, and get the retention terms in writing.

Sources

read 21 November 2026

Check it at the source.

  1. Signal: Signal is Serious About Protecting Your Privacy (Meredith Whittaker, 20 November 2026: what Signal blocks, what it cannot, and what it asks of Apple, Google and Microsoft)
  2. Cybernews, from Reuters: Signal blocks screenshots to stop AI agents reading chats (the Muse report and Meta’s response, as republished 20 November 2026)
  3. BRev Digital: an AI assistant kept reaching for company data after being told no (our check of the agents that run while nobody is watching)
  4. BRev Digital: how to stop an AI agent that is already running (our kill-switch guide, tested six ways)

A question?

Know what your assistants can see?

Tell us which AI assistants and browser tools your team runs. We will list what each one can see and sign in to, narrow the profiles, and check that your sensitive apps are really excluded.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com