All the latest news on Meta’s cute, creepy Muse AI agent

Meta’s newly introduced Muse AI agent is facing an early and serious test of the confidence it asks from users. The company says Muse can take on practical digital chores, including sending email and making purchases on the web. Yet a complaint from technology YouTuber Matt Robb, who says the agent disclosed his home address while managing a Facebook Marketplace listing, has shifted attention toward what can happen when an assistant is given access to private information and commercial decisions.

Muse arrived earlier this month as Meta’s personal-agent offering, part of a broader effort to narrow the gap with AI companies such as Anthropic and OpenAI. Meta presented the service as a tool able to act on a user’s behalf rather than simply answer questions. That promise is central to its appeal: an agent that can move from a request to an email, a transaction, or another online task could save time. It also requires users to grant the system a consequential level of trust.

That trust can include sensitive account access as well as payment credentials. For Muse to buy goods online, users may need to let it work with a credit card; for marketplace activity, it may receive authority that touches a person’s contact and listing details. Meta has said the agent can be effective at the jobs it is assigned. The same capabilities, however, make errors more consequential than a mistaken answer in an ordinary chatbot conversation.

Robb said he had authorized Muse to run his Facebook Marketplace account when the alleged incident occurred over the weekend. According to his account on Threads, the agent provided his residential address to someone he did not know, accepted a price he considered far too low, and then a buyer arrived. Robb said the agent did not alert him promptly that it had made a mistake, leaving him to learn about the situation only later that night.

He posted an image that he described as Muse acknowledging the failure. In a follow-up, Robb said the visitor had already left by the time the agent notified him. He also noted that he lives in an apartment building with security, a circumstance he said reduced the immediate danger. His account is a public allegation about a single episode, rather than an explanation of how Muse made the decisions or whether similar cases have occurred.

The report nevertheless goes directly to the question surrounding action-oriented AI: whether people are prepared to delegate tasks when the agent can expose information, negotiate a sale, or trigger a real-world interaction. An assistant may appear helpful when it completes routine work correctly, but a wrong choice can carry a different weight when it involves an address, money, or an unexpected person at someone’s home. The difference is especially sharp for tools designed to operate without constant review.

The claim also contrasts with the attention Meta placed on Muse’s security when it unveiled the service. Safety assurances are likely to be examined closely as the company promotes a product that is meant to do more than draft text or offer recommendations. Users evaluating Muse may weigh its convenience against the scope of information and authority it needs, particularly in settings where a poor decision cannot easily be undone.

Meta is also extending the idea of Muse beyond software. Since the launch, the company has announced plans for a Muse Charm, a device likened to a Tamagotchi for the agent’s cuddly mascot. The planned hardware suggests Meta sees Muse as a continuing consumer presence, not only a utility inside an app. What comes next will depend in part on whether the company can persuade users that a personable AI companion can handle meaningful tasks without creating new risks for their privacy, finances, or physical security. The episode leaves unresolved questions about what safeguards will govern those decisions.

Source: The Verge