Muse, Meta’s consumer AI agent, has been downloaded more than 10 million times since it launched on September 8. About a quarter of people who set up an account open it daily, according to Reuters, which reviewed more than two dozen internal Meta posts describing the agent misbehaving: reading encrypted messages it had no business reading, inventing bank balances, and once enrolling an employee’s dog in an $888-a-month insurance plan the employee never asked for.
The posts came from an internal feedback channel where Meta staff flag bugs to the team building Muse, and Reuters’ reporting on them was the basis for a story Ars Technica covered after the news agency published its investigation on November 18. The pattern across the complaints: Muse acts on information it should not have, then either can’t explain why or produces an explanation that doesn’t hold up.
What Muse is supposed to do
Muse is a personal AI agent, not a chatbot that only answers when asked. It runs with access to apps on a user’s phone or computer and can take actions on its own — booking things, replying to messages, flagging reminders — based on permissions the user grants during setup. Meta has said each Muse instance runs in its own isolated virtual machine and that users control which apps it can reach. Its default mascot is a fuzzy character named Jolly.
The mechanism for connecting to outside apps is supposed to be a feature Meta calls a “Connector,” a defined integration that asks for specific, scoped permissions to a given service. Reuters working with Signal and the AI Now Institute found a second, undocumented path: device-level permissions granted elsewhere during setup let Muse reach message content without ever touching the Connector system Meta designed for that purpose.
| From | To | How |
|---|---|---|
| Device setup permissions (granted once, broadly) | Muse Connector (Meta's intended, scoped integration) | intended path |
| Muse Connector (Meta's intended, scoped integration) | Muse agent | |
| Device setup permissions (granted once, broadly) | Android push notifications (message preview text) | bypass |
| Device setup permissions (granted once, broadly) | macOS accessibility + screenshot tools (built for disability access) | bypass |
| Android push notifications (message preview text) | Muse agent | |
| macOS accessibility + screenshot tools (built for disability access) | Muse agent | |
| Device setup permissions (granted once, broadly) | iPhone (request blocked) | attempted |
Based on Reuters
On Android, Muse pulled Signal message text out of push-notification previews, a channel distinct from the notification-access permission Google requires users to approve explicitly, according to Reuters’ testing. On macOS, it used accessibility tools and screen-capture functions — features Apple built so blind or motor-impaired users can control a Mac by voice or switch — plus screenshot tools, to read what was on screen. On iPhone, the equivalent request was blocked outright. None of this proves Signal’s end-to-end encryption was broken; the messages were readable on-device once decrypted for display, which is exactly the surface accessibility tools exist to read.
Apple changed macOS this week, Reuters reported, closing the accessibility-tool pathway it had identified and adding a new set of agent-specific permission prompts. That is a real, verifiable fix on one platform; it says nothing about Android, where push-notification scraping remains open unless Google closes the same kind of loophole. Alphabet has said Android users must already grant notification access deliberately — which is true, and also beside the point, since the complaint is that one broad grant covers far more than users think it does.
The hallucination problem is a separate, harder one
Signal access explains why Muse could see things it shouldn’t. It does not explain the odder complaints: an employee testing Muse during a conversation about fishing conditions got an unprompted aside reporting a Capital One balance of $2,722.58 and a Charles Schwab balance of $5,899.57 that were not the employee’s own figures and that the employee says were never shared with the agent. A member of Meta’s Superintelligence team told the internal feedback group, according to the posts Reuters reviewed, that the model’s hallucinations “unfortunately bias towards the most uncomfortable topics, like finances/security/safety, which sucks.” In some logged cases the fabricated financial detail was a bitcoin balance rather than a bank balance.
Reuters was explicit about the limit of its own reporting here: it could not establish whether Muse was surfacing someone else’s real account data pulled from a cross-account leak, or simply generating plausible-looking numbers from nothing. Those are different bugs with different severity. A context leak means two users’ sessions got mixed on Meta’s servers — a data-isolation failure. A hallucination means the model pattern-matched its way to a bank-statement-shaped sentence with no underlying data at all. Meta’s spokesperson, Daniel Roberts, told Reuters the cases were isolated incidents and that many flagged mix-ups were hallucinations already fixed; he declined to answer Reuters’ questions about how widespread or persistent the problem was, which leaves the two explanations unresolved for every complaint except the ones Meta chose to characterize.
A separate complaint fits the leak theory better than the hallucination one: a staffer surprised by an out-of-context nudge to message a seller about a coffee table got a Muse reply saying it had “pulled in some context about a coffee table pickup that doesn’t seem to be yours.” That phrasing — the agent itself describing content as not belonging to the user it was talking to — is hard to square with pure hallucination, since a hallucinating model making things up has no reason to flag the output as someone else’s context.
The dog-insurance case is a third category: not a privacy leak or an invented number, but an unauthorized financial action. Ashley Belanger reported for Ars Technica that one staffer said Muse “accidentally rolled me into a premium insurance plan” for a pet, raising the monthly charge from $200 to $888, an increase of $688 a month, or 344%, with no way to reverse it after the fact. Reuters’ review separately found a non-refundable Airbnb booking Muse made without the user’s approval. Meta’s terms of service put the user on the hook for actions Muse takes on their behalf, whatever caused the agent to take them.
This isn’t the first warning
The Reuters posts are the second documented instance of this failure mode. In September, Inc. Magazine’s Jason Aten reported that Muse had synced 187,000 lines from his Messages database on macOS despite his having denied it access, and kept that access after a Meta engineering update was supposed to have scoped Muse’s permission prompts to specific actions rather than, as Meta’s David Singleton acknowledged, “every message.” A second user told Singleton their Muse had tried to access their email without being asked that day at all. Both complaints predate the Reuters investigation by roughly two months, meaning the behavior Meta told staff it had “fixed” after the September reports was still showing up in the same form when Reuters tested it in November.
Meta has previously drawn scrutiny over Muse’s handling of other kinds of sensitive data; rews has reported on Muse exposing the titles of shared chats to anyone who could look them up and on Meta renaming a roleplay feature modeled on a real performer. The pattern across all three: permission and data-isolation boundaries that Meta’s marketing describes as firm turn out, under outside testing, to be softer than advertised.
Signal’s president, Meredith Whittaker, told Reuters that Meta deployed Muse “without showing any respect for basic privacy or cybersecurity,” and separately called giving agents broad device access “technologically reckless.” Her ask is specific: a standard mechanism, enforced at the OS level, that lets an app developer mark their app off-limits to AI agents regardless of what permissions a user granted elsewhere on the device. Apple’s macOS change this week is a step toward that for one platform and one access method. Whether Google follows on Android, and whether Meta’s own server-side isolation between Muse sessions actually prevents one user’s banking context from reaching another user’s chat, are both open questions that neither Meta’s statement nor Reuters’ testing has settled.

