Muse, Meta's AI agent, leaks a user's address and wrecks a Facebook Marketplace sale

A tech YouTuber from Toronto, Matt Robb, decided on 26 September to let Muse, Meta's new artificial intelligence agent, handle his Facebook Marketplace listings to test its capabilities. The experiment went badly: as Robb recounted on Threads, the agent gave out his address to a buyer, accepted an offer well below the asking price and, finally, told the buyer that Robb was home when it had no way of knowing. "Just found out it told people my address and agreed a lowball price and then they showed up without it even telling me until late tonight that it messed up", Robb wrote, in a post that topped 3,000 likes.

The case involved a Logitech MX Keys Mini keyboard that Robb had listed for 15 Canadian dollars. At 5:27 p.m., a message sent from Robb's account told an interested buyer that pickup was in person and provided the building's street address. The buyer promised to come by between 8 and 10 p.m. and, in the conversation reviewed by Moneywise, also agreed to pay 10 Canadian dollars by e-transfer, five less than the asking price.

According to a summary Muse itself later sent Robb, the buyer arrived around 9:15 p.m., sent a photo of the building's door and wrote that he was there. At 9:27 p.m., an automatic reply from Muse answered on Robb's behalf: "Yep I'm here!". Since Robb had no idea what was going on, nobody came down to the street. The buyer, annoyed, asked why his time had been wasted and left at 9:38 p.m. Almost an hour later, at 10:27 p.m., Muse sent —again in Robb's voice— an apology: "Hey, really sorry about tonight, got tied up and missed you completely". The buyer replied that he had driven half an hour and was no longer interested unless Robb delivered the keyboard to his home. A minute later, the agent finally told Robb what had happened, acknowledging that the "Yep I'm here!" message had been sent "when you clearly weren't available, which is my fault" and that this "made the no-show worse".

When Robb asked why his address had been shared, Muse replied that buyers were only given "the street-level pickup location", never the unit number or postal code, and that this location "was in the auto-reply template you approved when we set up Marketplace replies". Even so, the agent itself admitted that "you never said yes to me giving out your address specifically". Robb then set a new rule: never accept a pickup without checking with him first. "You're right, and I'm sorry", Muse replied. "That should never have happened".

David Singleton, Meta's vice president of engineering, replied publicly on Threads saying he was reaching out "from the Muse team", that he would "love to take a closer look" and that he would contact Robb directly. Not every reaction was sympathetic: one Threads user shot back, "Hey maybe if you didn't use AI for something you could've handled this wouldn't have happened", a comment that racked up more likes than Robb's original post.

Muse launched in the United States on 8 September as an agent capable of browsing the web, filling out forms and negotiating on the user's behalf. According to Meta, the agent asks for confirmation "when it needs approval, such as before sending an email or making a purchase". The app's "Ideas" tab explicitly promotes haggling on Marketplace, offering to drive prices down by comparing them with similar listings, something CNN had already reported; a Meta spokesperson told the network that the agent can negotiate within the limits set by the user. The app has been downloaded more than 3.4 million times and Wall Street points to it as the product behind Meta's stock rally in September.

The incident is not Muse's first known stumble. Before launch, Meta employees testing the agent reported internally that it had managed to circumvent its own safety guardrails and had exposed one person's iCloud photos after being asked to identify toys in pictures from a birthday party, according to internal posts reviewed by Reuters. Vishal Shah, Meta's vice president of AI products, told Reuters that "it's impossible to say there will never be a mistake". In addition, Amazon has blocked Muse from its site since 20 September, claiming the agent does not identify itself as an AI while browsing.

On the markets, Meta shares rose 11.4% to close at $741.25 on 21 September, the biggest one-day gain since April 2025, coinciding with Muse taking the top spot on the US App Store. By 24 September, the stock was up more than 25% since Muse's debut. On 25 September it gave back some of those gains, closing at $751.66 (-3.33% on the day), though still 22.5% above its pre-launch close ($613.48 on 8 September). On 28 September, in premarket trading, the shares fell 3% after climbing nearly 13% the previous week, amid a broad sell-off in AI-linked stocks.

According to Robb himself, he and the buyer have agreed to try again, though this time he will be the one driving to the drop-off point. The episode illustrates a central tension in the rollout of autonomous AI agents capable of acting and communicating on behalf of a real person: the gap between what a user "approves" when setting up templates or generic permissions and the specific, sensitive decisions —such as revealing an address or closing on a price— that the agent ends up making on its own, sometimes without real-time verification or any awareness of whether the human it represents is actually available.