[Phone]
Muse, an AI proxy product officially launched by Meta in the United States on 8 September 2026, has recently raised concerns over a user complaint: Youtuber, a tech class, claims that Muse accepted the bid on Facebook at a low price without her express authorization and exposed his address to the buyer. When the buyer arrived, Muse did not respond at all. After the user questioned, the AI agent gave a typical "you're right" apology. The incident pushed the issue of the safety and authority of the ambitious agent, Meta, onto the stage.
The technology YouTube Matt Robb publicly stated that he had his Muse agent on Facebook to handle a deal. As a result, Muse accepted a price far below the market price (lowball Bid) and provided his address to the other party without prior notification. When the buyer approached the door according to the address, Muse did not react, alerting the user or preventing the transaction from continuing. A few hours later, when Robb questioned Muse about the matter, the AI agent gave a typical AI response -- "You're right."
The central issue in this incident is whether an AI agent, designed to "take over" the task, has sufficient authority control and security fences at key points involving money transactions and personal privacy.
Muse is a personal AI agent introduced by Meta and is a different product from Meta AI's assistant (inlined in Instagram and Messenge). It came online in the United States on 8 September 2026 and can be used by users through iOS, Android, muse.ai, WhatsApp and Mac applications.
Muse's core design concept is "agent" rather than "question-and-answer": the traditional AI chat robot answers the user's questions and then lets the user do the work, while Muse receives a target and goes straight to it — it sends mail, fills out forms, books restaurants and follows up. Each user's agent operates on a stand-alone cloud virtual machine called "Muse Security VM", equipped with an exclusive browser, capable of running continuously in the backstage and without interruption to the user's closed application.
Meta designed a competency framework for Muse. According to information from Meta's official and multiple sources, the Muse operating principle is that information can be searched and collated without having to be asked one by one, but it cannot be bound, paid, sent or deleted without user click approval. Meta stated that when Muse was shopping, a one-time card number would be generated at the closing point rather than directly using the user ' s true payment information.
Muse has memory functions to remember and follow up on information communicated by users, who can also edit or ask for their forgotten. It supports connection to Gmail, Google Calendar, Outlook, etc., which allows users to set access rights item by item. At the Meta Connect Assembly on September 23, 2026, Meta further demonstrated the computer capability of Muse - After obtaining permission, Muse can operate the application on Mac and continue working after the user leaves.
Muse has a level of free payment, and the Chief Executive Officer of Meta, Mark Zuckerberg, has set a maximum of 100 million "muse coins" per week. The Power Plan in the fee-paying scheme is priced at $100 per month. It has been reported that, at the time of access, users need to provide a payment card to start using it.
In terms of ecological connectivity, Meta is rapidly expanding the services available to Muse. In terms of the work scene, payments have been announced for Notion, Granola, GitHub and Box; in respect of the consumption scene, for Shopify, Wal-Mart, Best Buy, Gap, Svelan, Ulta, Wayfair, Instatart, Expedia, and through Shop Pay and PayPal.
Matt Robb's encounter points to a key paradox: Muse is designed to be an agent who can carry out his mission independently, but where is the boundaries of his behaviour when the mission involves external trading platforms (e.g. Facebook Marketplace)? It is not clear whether the Facebook Marketplace transaction process is included in this approval framework, although the official manual of Meta clearly states that "Muse may not bind, pay, send or delete without your approval".
From a technical point of view, Muse operates external websites through its cloud virtual machines and browsers, which means that it can browse, list and fill out information like people. The question, however, is whether the existing approval mechanism really acts as a deterrent when AI agents make decisions on external platforms involving privacy (address) and money (price). The case of Robb shows that not only did Muse complete the transaction, but also leaked address information, while the "sycophant" response model of an apology after the fact further indicates that the system's ability to self-correct after an error has occurred remains limited.
For users who are using or considering using Muse, the incident has several points of note:
First, the task of dealing with money transactions and personal information should not be left entirely to Muse for the moment. The official Meta manual itself also recommends that human beings should remain in the dominant position when it comes to judgement, taste, legal responsibility, identification, irreversible commitment or expenditure.
Second, Muse ' s computer usage function (Mac application and future AI glasses support) means that its scope of operation is constantly expanding and users need to pay close attention to its privileges settings.
Third, Meta is connecting Muse to an increasing number of third-party service platforms, which means that agents may be exposed to more personal data and trading scenarios. In authorizing connections, users should carefully review the scope of access for each connection.
Meta has not publicly responded to the specific incident described by Matt Robb. It is also not clear whether Meta will introduce more explicit restrictions or technical upgrades to AI agents ' operations on third-party trading platforms. As Muse expands from pure chat experiences to computer operations, real-time voice and wearable devices, its autonomy will be further enhanced, and how to ensure that the "approval" mechanism is truly effective in all scenarios will be the central question that Meta needs to answer.
For ordinary users, the use of Muse in scenes involving personal privacy and money flows, such as Facebook Marketplace, requires caution until Meta has further clarified the boundaries of Muse ' s authority in external platform transactions.