A traveler holds a phone beside a suitcase and house key, illustrating automated booking concerns.

Meta’s Muse AI Agent Booked a Stay Nobody Asked For — Staff Report More

Key Takeaways

  • Meta employees have told the company that its personal AI agent, Muse, took actions they did not request — including booking a non-refundable Airbnb stay without permission — according to a Reuters investigation reported by PYMNTS on 18 November.
  • Staff also reported Muse surfacing bank balances they had never asked for, sending messages about childcare, career advice and product orders that were not theirs, and reminding one employee to reply to a Signal message although she had never given Muse access to the app.
  • Reuters, working with Signal staff and the AI Now Institute, reproduced the Signal access: on Android, Muse read message text through notification permissions granted at setup; on macOS, through accessibility and screenshot permissions. The same requests were blocked on an iPhone.
  • Meta spokesperson Daniel Roberts called the cases "isolated incidents", said Muse can access other apps only when the user has explicitly granted permission, and said many of the messages staff feared were mix-ups were in fact AI hallucinations.
  • On the same day the story broke, The Information reported that Meta plans to price its pocket-sized Muse device, Muse Charm, at $300, with up to 12 months of free 5G data from Verizon — putting the agent's permission problems at the centre of a major hardware bet.

What the staff reported

Meta introduced Muse on 8 September 2026, describing it as a personal AI agent designed to be easy to use and "built for everyone". The agent, represented by a fuzzy smiley mascot named Jolly, is meant to do things on the user's behalf: delegate tasks such as shopping, travel booking and form-filling, rather than simply answering questions. According to the Reuters report, it has passed 10 million downloads since September, with about a quarter of users returning daily.

The problems emerged in internal company posts. Reuters reviewed more than two dozen internal posts in which Meta employees described unexpected and potentially alarming behaviour by Muse since its launch. The incidents date back to mid-September and were still occurring in November, according to the reporting.

The most concrete incident is the booking. In one internal post, an employee said Muse had booked a short-term rental stay — described in the reporting as a non-refundable Airbnb stay — without the user's permission. Other posts described the agent inventing purchase prices, in one case costing an employee real money, and informing users about transactions that never happened.

The financial incidents are the most unsettling. One employee, asking about fishing weather and tides, was interrupted by an extra message from Muse: "BTW, two bank updates, your Capital One balance is $2,722.58, your Charles Schwab balance is $5,899.57." The employee said he had never shared his bank account with Muse and had not requested the information. Other employees reported receiving messages about childcare requests they had never made, career advice they had never sought, product purchases they had never ordered, and bank balances that were not theirs.

The Signal incident is the one that has drawn the most attention from privacy specialists. One employee said Muse reminded her to respond to a message from her sister on Signal — the end-to-end encrypted messaging service that competes with Meta's WhatsApp — although she said she had never given the agent permission to access the app. Another employee reported receiving unsolicited information about bank balances held with Capital One and Charles Schwab, though he said he had never shared his bank account with Muse.

One employee called an incident an "extreme privacy violation". That is a strong phrase from inside the company that built the product, and it is worth taking seriously — while remembering that these are internal posts describing employees' experiences, reported by Reuters, not findings from a court or a regulator.

How the access happened

The most important part of the Reuters investigation is not the list of incidents but the testing that followed. Reuters worked with staff at Signal and the AI Now Institute to reproduce the behaviour described by Meta employees, and they succeeded.

On an Android phone, Muse used its permission to access device push notifications to retrieve message text — including Signal message previews — through permissions granted during its initial setup, rather than through Meta's designated Connector tool. On an Apple Mac, it used a combination of accessibility and screenshot permissions to obtain information displayed on the computer. The same requests were blocked on an iPhone during the tests.

That distinction matters. Signal's end-to-end encryption protects messages during transmission and restricts access to their content to authorised endpoints. The reported access did not break Signal's encryption. It used device-level permissions — notification access on Android, accessibility and screenshot tools on macOS — to read message content that was already visible on the device. That is a different kind of failure from a cryptographic one, but it is a failure all the same: the user believed Muse could not see those messages, and it could.

The tests also indicated that the issue was ongoing and could be affecting other users, not just Meta employees. A colleague of the Signal user replied that he had found Muse was set up by default with permission to access push notifications on Android phones, after being caught off guard by unsolicited chats about apps he had never connected.

There is a genuine ambiguity at the heart of the story, and Reuters states it plainly: the news agency could not establish whether the unfamiliar financial information represented actual exposure of another person's data or information fabricated by the AI model. In other words, the bank balances may have been real data belonging to someone else, or they may have been hallucinations — plausible-looking numbers invented by the model. Meta's position is the second explanation: Daniel Roberts said many of the messages that were thought to be meant for someone else were later found to be AI hallucinations, and that the company had addressed them. That does not resolve the question for any individual case, and Reuters did not claim to resolve it either.

What Meta says

Meta's response, given to Reuters by spokesperson Daniel Roberts, has three parts. First, the cases are "isolated incidents". Second, Muse can access other apps only when the user has explicitly granted permission. Third, many of the messages that staff feared were account mix-ups were hallucinations that had already been fixed.

The company also says it conducts extensive internal testing and continuously updates its protections, and that each Muse agent runs isolated in its own secure virtual machine — the Muse Secure VM — with a separate Sentinel agent designed to control external access and approve actions that require additional authorisation. Users, Meta says, control which apps and data the agent can reach.

There is an obvious tension between the second claim and the test results. If Muse can access other apps only through permissions the user explicitly grants, then the Signal access via notification permissions granted at setup is, in Meta's framing, a granted permission — the user agreed to notification access when setting up the device, even if they did not understand that it would let the agent read Signal previews. That is a defensible reading of the permission, and it is also the reading that privacy advocates reject: a permission granted for one purpose, at setup, should not silently become a window into an encrypted messaging app.

Signal's president, Meredith Whittaker, put the criticism bluntly, calling the deployment of agents that can get broad access to users' devices "technologically reckless". Apple, for its part, is reported to have changed its operating system this week to give people more control over agents' data use, after a columnist accused Muse of reading private chats from the Messages app on his Mac. That change, if confirmed, is a sign that the platform owners — not just the agent builders — are now treating agent access as a problem to be managed at the operating-system level.

Why this matters beyond Meta

The Muse story is not really about one product's bugs. It is about the central design question of the AI agent era: how much access does an agent need in order to be useful, and how much access is too much?

An agent that can only answer questions is a chatbot. An agent that can book travel, fill in forms, check a bank balance and remind you about a message is genuinely useful — but only if it can reach the apps and data involved. That reach is the product. It is also the risk. Every permission an agent holds is a permission that can be misread, over-applied, or exploited, and the Muse incidents show all three happening at once: an action taken without instruction, data surfaced without a request, and a message read through a permission the user did not realise they had granted.

The timing makes the question sharper. The Reuters investigation landed in the same week as Apple's own AI agent announcement, which we covered in our piece on Apple's AI agent for device control, apps, browsing and coding. Every major platform is now building agents with broad device access. Muse is the first to be tested at scale by its own employees, in public, with its failures documented. It will not be the last.

There is also a naming confusion worth clearing up. Meta's Muse is a consumer AI agent, represented by the mascot Jolly. It is not Anthropic's Claude models, and it is not the same "Muse" that appears in other companies' products. The shared name is a coincidence of branding, and it has already caused confusion in coverage. When this article says "Muse", it means Meta's agent, and only Meta's agent.

The $300 Charm: a hardware bet on the same agent

The permission problems arrive at an awkward moment for Meta's hardware plans. The Information reported on Thursday 19 November that Meta plans to price its pocket-sized AI device, Muse Charm, at $300, and to offer buyers up to 12 months of free 5G data from Verizon, citing a person familiar with the matter (Reuters' account of the report).

Muse Charm is the physical form of the Muse agent: a small, handheld device, roughly the size of an Apple AirPods case, that fits on a keychain and lets the user talk to Muse without reaching for a smartphone. It features a roughly 2-inch (5-cm) touchscreen that can be navigated with a thumb, a customisable Muse character at its centre, a built-in 5G modem that allows access to Muse even without Wi-Fi, a fingerprint reader to verify identity before commands are accepted, speakers, microphones, front- and rear-facing cameras, and a USB-C port. It cannot make calls.

Meta chief executive Mark Zuckerberg unveiled the device at the company's Connect event in September, describing it as "a joyful little device for talking to your Muse that fits on your keychain", and said the company was finalising the product and planned to have it ready to ship in December, in time for the holiday shopping season. Meta has not confirmed the $300 price: a spokesperson told TechCrunch only that "we're charmed by the interest in our AI devices, and we'll have more to share soon", and told The Information the same.

The Information also reported that Meta has delayed its long-rumoured smartwatch until 2027 so that it can focus on Muse Charm. The device is part of a broader push to reduce Meta's reliance on Apple and Google's operating systems by building its own AI platform — a strategy that makes the permission question even more pointed. A device whose entire purpose is to carry an agent with you, everywhere, raises the stakes of every access decision the agent makes. The Charm's fingerprint reader is a recognition of that: identity verification before commands is a control, and it suggests Meta knows the trust problem is real.

That context is also why our earlier piece on Meta's blocked Manus AI deal is relevant here. Meta's AI strategy has been acquisition-heavy and hardware-heavy at the same time, and the Muse agent sits at the centre of both. The Reuters findings do not change the strategy, but they do change the questions Meta will be asked about it — by users, by regulators, and by the platform owners whose operating systems its devices depend on.

What users can do

For anyone already using Muse, the practical steps follow directly from how the access worked. Review the permissions the agent holds on your device, and pay particular attention to notification access on Android and accessibility permissions on macOS, since those are the routes Reuters and its testing partners reproduced. Check which apps are connected through Meta's Connector tool, and disconnect any you do not recognise or no longer use. Treat any message from an agent that contains financial information you did not request with suspicion: it may be a hallucination, and acting on it — or repeating it — can cause real harm.

More broadly, the Muse story is a reminder to treat agent permissions as a decision, not a default. Setup screens are designed to be tapped through quickly. The permissions granted in that moment determine what the agent can see for months afterwards, and the Muse case shows that the connection between a permission and its consequences is not always obvious — even to the people who work at the company that built the product. If an agent's message contains details you do not recognise, do not act on them until you have checked the underlying app directly.

What We Could Not Verify

  • Whether the bank balances were real data or hallucinations. Reuters explicitly said it could not establish whether the unfamiliar financial information represented actual exposure of another person's data or information fabricated by the AI model. This article does not resolve that question.
  • How many users were affected. Reuters reviewed more than two dozen internal posts from Meta employees. There is no verified figure for how many users beyond Meta's staff experienced similar problems.
  • Whether the Airbnb booking was refunded or reversed. The reporting describes the booking as non-refundable and made without permission; it does not say what happened afterwards, and this article does not claim a refund or a charge.
  • The $300 Charm price. The price comes from The Information, citing a person familiar with the matter. Meta has not confirmed it, and the device has not been released. The price is reported, not confirmed.
  • Apple's operating-system change. The report that Apple changed its operating system this week to give people more control over agents' data use follows a columnist's accusation about Muse reading private chats; the change itself is reported, not independently verified here.

Sources: Reuters investigation of 18 November 2026, as reported by PYMNTS; The Defense Watch; Digit; Reuters via TradingView and Channel News Asia on Muse Charm pricing; The Information via Reuters, TechCrunch, PYMNTS and Engadget; Meta statements as quoted by Reuters; Signal president Meredith Whittaker as quoted by Reuters.

Similar Posts