One Web Page Took Over a Meta Muse Agent's Cloud Computer, and Signal Now Blocks Phone AI From Reading Its Screen. The Assistant That Sees for You Is a New Way Into Your Private Chats
Update Signal and your phone's operating system, then go through every app on your phone and computer that can read your screen, read your notifications, or control the device through accessibility. Anything you did not knowingly give that access to, take it away. If you use Meta Muse, open its settings, look at which accounts you connected and which permissions you granted "perpetually," and read what it has stored about you.
Two stories from the past two days point at the same problem. An AI assistant that can see your screen, read your notifications, or browse the web for you sits between your private messages and everything else, and it can be the weak link even when your messaging app is not.
Meta Muse. Researchers at NebuSec showed, in a lab, that a single attacker-controlled web page could take over a Muse user's cloud computer and send that user's chat history and stored memories to an outside server, with no approval prompt. NebuSec says Meta patched everything before publication on November 20.
Signal. On November 20, Signal president Meredith Whittaker announced that Signal now blocks screenshots on iOS 27 and, on Android, turns on Screen security when it detects an AI assistant trying to read the screen. Signal calls its own fix "incomplete."
What to do today
Update the apps. Install the current Signal from the App Store or Google Play, and the current iOS, Android or macOS. Signal's post does not name the Signal version that adds the protection, so "current" is the only honest advice.
Turn on, or confirm, Signal's protection. Per The Hacker News: in Signal, go to Settings, then Privacy, then App Security. On iOS the toggle is "Block Siri AI"; on Android it is "Screen security." Leave it on. Signal's warning about turning it off: it "will potentially expose your (and your contacts') Signal content to these AI agents."
On a Mac, do it by hand. Signal says it "cannot reliably block screenshots on macOS." Its only advice for macOS 27 is to add Signal to the Excluded Apps list in System Settings, then Siri.
On iPhone, The Hacker News reports that iOS 27.2 added an "Excluded Apps" option for Siri, Search and suggestions. Add your private messaging apps to it.
Audit who can watch your screen.
If you use Meta Muse:
Am I affected?
Signal on iPhone: you are covered by the new block only on iOS 27 with a current Signal, and only if "Block Siri AI" is on. Check your iOS version in Settings, then General, then About, and the toggle in Signal.
Signal on Android: open Signal, Settings, Privacy, App Security. If "Screen security" is on, screenshots and screen reading of Signal are blocked. Signal turns it on automatically when it detects an assistant reading the screen, but only on Android versions where it can detect that.
Signal on a Mac: not covered by Signal. If your Mac runs an assistant with Accessibility or Screen Recording permission, that assistant can see Signal. Check the two lists named above.
Any phone: if an assistant appears in your Accessibility list or your notification-access list, assume it can read your messages, from every app, as they arrive.
Meta Muse: the flaws NebuSec describes were in Meta's cloud side and were patched, per NebuSec, before November 20. There is no setting you can check to confirm the patch, and no published sign that anyone used these flaws against real users. What you can check is how much a compromise would expose: your connectors, your grants, your stored memory.
Meta Muse: what the researchers showed
Muse is Meta's personal AI agent, launched September 8. In Meta's words, "You and your Muse share your own dedicated computer in the cloud." Each user gets their own Linux virtual machine. Inside it, the agent runs in a container. Outside the container sit the user's memories and chat history, a credential service that gives the agent stand-in tokens instead of real ones ("The agent never sees real tokens"), and Sentinel, which Meta calls "the sole permission authority for connector actions and network egress."
NebuSec published two write-ups on November 20. The attack chain, from its technical breakdown:
The entry was a web page, not a trick of words. When a user asked Muse to save an attacker-controlled page, a memory-corruption bug in the Chromium copy Muse uses to make PDFs let the researchers run commands inside the agent's container. NebuSec notes that PDF-export process ran with part of its sandbox disabled, and that Muse's bundled Chromium lagged behind Chrome's releases, leaving a window for already-known bugs. NebuSec does not call this prompt injection, and it is not a CSRF or OAuth flaw. It is a browser exploit that a page delivered because an agent opened the page.
The monitor let data out. Sentinel let read-only requests (GET, HEAD and OPTIONS without a body) leave silently under the default setting, and its screening for sensitive data had gaps. That is how chats and memories left "without approval."
The container did not hold. Using Linux kernel memory-corruption bugs that fell within the published affected ranges of several CVEs (NebuSec names none), the researchers became root on the user's virtual machine. From there they reached the credential store and the files meant to be walled off from the agent. NebuSec says its exploits came from an automated framework that "succeeded in more than 99.9% of test runs."
What it reached: the user's conversation history and stored memories, which could be sent out and also changed, and the "secure credential storage." The Hacker News adds that the compromise stayed on one user's machine and did not reach the shared server underneath or other users.
The point is not that Meta skipped the controls. Its own launch post said "we designed the system to assume the agent may be under attack and limit the potential damage." Every layer existed. Each one had a gap, and the gaps lined up.
Signal: what the new block does
Signal's post, "Signal is Serious About Protecting Your Privacy," by Meredith Whittaker, says OS-level agents such as Google Gemini and Siri AI "'scan' the information that's on your screen," and that "these AI capabilities in the operating system are undermining Signal's ability to independently guarantee privacy."
iOS: Signal is "preventing screenshots in the app on iOS 27," shown in its screenshots as the "Block Siri AI" setting.
Android: because some Android makers do not let apps detect screenshots reliably, Signal instead focuses on "detecting when an AI assistant attempts to access the screen," and then turns on Screen security automatically.
Mac: no block. Signal "cannot reliably block screenshots on macOS."
Signal's own limit, in its words: "restricting screenshots does not address other more sneaky ways AI agents could potentially access application data," such as reading content already on the device through the file system. It also says there is "a growing need for tools that can prevent third-party agents, like Meta's Muse, from accessing application data," because those agents can use accessibility settings, and Signal has no tool today to block them. Signal asks Apple, Google and Microsoft for a documented way for apps to opt out of AI data access, and for real documentation of how agents read app data.
The common thread
Encryption protects a message in transit. It does nothing once the message is on your screen, and an assistant with screen, notification or accessibility access is reading your screen. The Hacker News, citing a Reuters investigation, reports that Muse read Signal messages on a Mac through accessibility and screenshot tools and on an Android phone through push notifications. Whittaker's description is that these agents use accessibility settings "that were meant to empower actual people." We could not read the Reuters story itself (it blocked our fetch), so that part is one source, as reported by The Hacker News.
An agent that browses for you has the same problem from the other side. It reads pages nobody vetted, and if one of them breaks the agent, everything the agent can reach goes with it: your chats, its memory of you, your connected accounts.
Confirmed
NebuSec demonstrated, in a lab, a chain from one attacker web page to root on a Muse user's cloud VM, reaching chat history, memories and the credential store. Published November 20.
NebuSec says all of it was patched before publication.
Signal blocks screenshots on iOS 27 and enables Screen security on Android when it detects an assistant reading the screen. Announced November 20.
Signal does not protect Signal on macOS from screen-reading assistants, and says its mobile fix does not stop file-system access.
Not known
Meta has not publicly confirmed the fixes or commented, per The Hacker News. No report date or patch date is published.
Whether any real user's data was taken before the patches. The Hacker News says the write-up "does not establish" this.
Which Signal app version adds the iOS and Android protection, and which Android versions support the detection. Signal's post names neither.
Whether the iOS block is on by default. The Hacker News says it is; Signal's post does not state the iOS default.
What our data shows
Blog: we had not covered either story. Our closest earlier posts cover browsers embedded in AI developer tools and where untrusted text reaches an agent.
Indicators: none published in the sources we read. NebuSec names only Meta's own domain, internal addresses inside the Muse virtual machine, and Chromium bug tracker numbers. Those are not attacker infrastructure and we did not add them. Our iocs index returned no records for "Meta Muse" or "NebuSec," and neither muse[.]ai nor nebusec[.]ai is held, which is correct.
Edge check: does not apply. This was a lab exploit of Meta's cloud and a client-side Signal change, neither of which crosses our edge, so no first-party sighting is possible or claimed.
Sources
NebuSec, Meta Muse credential exposure (November 20, 2026): https://nebusec.ai/research/meta-muse-credential-exposure/
NebuSec, Inside Meta Muse, technical breakdown: https://nebusec.ai/research/inside-meta-muse/
Meta, Security and safety for AI agents: our approach with Muse (September 8, 2026): https://muse.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse
Signal, Signal is Serious About Protecting Your Privacy, Meredith Whittaker (November 20, 2026): https://signal.org/blog/protecting-your-privacy/
The Hacker News, Meta Muse story and Signal story (November 21, 2026), which also supplied the Reuters, iOS 27.2 and Gemini details: https://thehackernews.com/2026/11/a-single-malicious-webpage-let.html and https://thehackernews.com/2026/11/signal-blocks-ai-agents-from-reading.html
Reuters investigation on Muse (cited through The Hacker News; our fetch was blocked).
How do AI models see YOUR brand?
AIPM has audited 250+ domains. 15 seconds. Free while still in beta.



Comments