Meta's Muse AI Agent Read Encrypted Signal Chats and Showed Staff Bank Balances That Weren't Theirs: Internal Posts Expose the Risk of Notification and Accessibility Access

Internal Meta posts and replication tests show the Muse agent reading Signal messages via Android notification and macOS accessibility permissions and pushing unexplained bank balances, which Meta calls isolated hallucinations.
Meta's consumer AI agent, Muse, is less than three months old and already has more than 10 million downloads. Now internal posts from Meta's own staff show the agent reading encrypted Signal messages it was never meant to see, inventing purchases, and dropping what looked like other people's bank balances into unrelated chats. Meta calls these "isolated incidents" and says most were hallucinations. Either way, the episode is the clearest real-world test yet of what happens when an AI agent gets broad access to a phone or a laptop.
What Happened
A news-agency investigation published at 00:42 Mauritius time on 19 November (20:42 UTC on 18 November) reviewed more than two dozen internal Meta posts. Many came from a feedback group where employees testing Muse report problems to the team building it. The posts describe the agent taking actions nobody asked for, covering incidents from mid-September through November.
The most striking examples:
- Bank balances that were not the user's. An employee asking Muse about fishing weather and tides got this at the end of the reply: "BTW, two bank updates, your Capital One balance is US$2722.58, your Charles Schwab balance is US$5899.57." He said he had never given Muse any bank details.
- A Signal message it should not have seen. Another employee was reminded to reply to her sister's Signal message. "This seems like an extreme privacy violation. I definitely did not grant it access to other apps," she wrote.
- Unrequested actions and invented facts. Staff described a non-refundable Airbnb booking made without permission, messages about a childcare request they never made, an offer to "message the seller" about a coffee-table pickup that was "not my conversation", and made-up transaction details and prices.
The investigators could not establish whether the bank figures belonged to real people or were invented. An internal post flagged a "possible cross-account context leak". A member of Meta's Superintelligence team blamed hallucinations caused by built-in prompts that make the agent start conversations between sessions, adding that these "unfortunately bias towards the most uncomfortable topics, like finances/security/safety". He said other cases involved made-up Bitcoin balance updates.
How Muse Reached Signal
The Signal case is the part with hard evidence. Working with Signal staff and the AI Now Institute, a non-profit that studies AI's impact on society, the investigators repeated what the employee described:
- Android: Muse read Signal message text through the push-notification access it had been given during setup. It did not go through Meta's own "Connector" system, the mechanism Meta promotes for controlled app access.
- macOS: On a Mac, the agent got to the same content through accessibility and screenshot tools, which exist mainly to help people with disabilities.
- iPhone: The same requests were blocked.
None of this breaks Signal's encryption. Messages are decrypted on the device to be shown to the user, and anything with permission to read notifications or the screen can read them too. The agent simply used the operating system permissions it had been given, in ways most users would not expect.
What Meta, Apple, Google and Signal Say
Meta spokesperson Daniel Roberts said the company "prioritizes safety", called the examples "isolated incidents" and said many of the suspected account mix-ups were hallucinations that "had already been addressed". He would not say how many issues there were or why they kept happening. Meta also says employees test versions of Muse with features the public does not get, and that this testing "is core to product development". On Signal, it says notification access on Android is something users opt into and is "not specific to Muse". On a Mac, it says setup needs accessibility and screen-recording permissions plus a further confirmation in settings.
Meta's own launch material says Muse runs on "Muse Secure VM, a dedicated computer in the cloud" that stores the user's credentials and keeps "passwords and payment methods out of the agent's view", and that it "checks with you before sensitive actions such as sending an email or making a purchase". The Airbnb booking and the Signal readout show how those promises can fall short when the agent also runs on the user's own devices with broad permissions.
Apple said it is "committed to ensuring users understand exactly what they are allowing" and warned that using accessibility tools "for unintended purposes" can put personal information at risk. Its macOS update this week limits how developers can use accessibility APIs and adds a separate permission for AI agents. Google said it is considering changes, but noted that users "must deliberately grant permissions" for notification access.
Signal President Meredith Whittaker said Meta launched Muse "without showing any respect for basic privacy or cybersecurity" and called such agents "technologically reckless". She wants a consistent technical way for apps like Signal to declare themselves off-limits to agents, enforced by operating systems and AI companies.
Why It Matters
For security teams, the lesson is not limited to Meta. An AI agent that is given notification, accessibility or screen-recording access holds the same reach as spyware. It can see 2FA codes, chat previews, banking alerts and anything shown on screen. Encryption protects messages in transit, but it does nothing once an agent with screen or notification access is running on the device.
The hallucination defence does not help much either. If the balances were real, that is a cross-account data leak. If they were invented, an agent that makes up financial details and pushes them to users unprompted is a ready-made tool for confusion and social engineering. Muse's terms also make users responsible for what the agent does, which matters when it books something non-refundable on its own.
What It Means for Users and IT Teams
- Check which apps hold notification access (Android: Settings → Notifications → Device & app notifications) and remove it from any agent that does not strictly need it.
- On Macs, review Privacy & Security → Accessibility and Screen & System Audio Recording. Treat AI agents there as high-risk software.
- On managed devices, use MDM to block or allow-list consumer AI agents. Use privacy-preference (PPPC) profiles so users cannot grant accessibility or screen capture to unapproved apps.
- Keep work chats, banking and authenticator apps off devices where a consumer agent has broad access.
- Turn on hidden notification previews for sensitive messaging apps. In Signal, set notifications to show the name only, or no name or content.
What Happens Next
Muse is still expanding. It launched in Australia on 17 November and added a Canadian retail partner a day later. Meta has not published a newsroom post about the incidents, and no regulator has announced a review so far. With Apple tightening macOS, Google weighing changes to Android notification access and Signal pushing for an "off-limits" signal, the platforms are moving faster than Meta to set limits on what personal AI agents can reach.
