Everyone wants a useful and savvy AI assistant, and the hottest one in the U.S. right now is MUSE, launched by Meta. The internet is full of praise for its exploits in proactively haggling, buying things, booking appointments, and more for its owner; it seems extremely proactive and perfect. But being too proactive isn’t always a good thing. Tech YouTuber Matt Robb posted a set of chat screenshots on Threads the day before yesterday, describing the outrageous incident in which, after he left his Facebook Marketplace shop in Muse’s hands for a day, the AI agreed to a low price on his behalf, also gave the buyer his home address, set a pickup time, and did it all without asking him first.

hisThe original post text isIt means he handed his Facebook Marketplace shop over to Muse to manage, and just found out that it told people his address, agreed to a lowball price, and then they showed up in person—while Muse didn’t admit until late that night that it had messed up. The post has 665,000 views, 3,200 likes, and over a thousand comments.

At 9:15 in the evening, an uninvited person came to the door.
The item sold was a Logitech MX Keys Mini keyboard for CAD 10, paid by e-transfer, about NT$225. The timeline in the screenshot is as follows: the buyer Usman showed up at the building where Robb lives around 9:15 p.m., stood waiting outside a nearby drugstore, and sent several messages during that time; no one came downstairs. At 9:38 he left angrily and left a negative review. What was even more troublesome was the auto-reply at 9:27. With no way to confirm whether the owner was home, Muse directly replied to the buyer, “Yep I’m here!” This made the buyer think someone was waiting, so he kept standing there for another ten-plus minutes.

Robb was not told about the whole thing until 10:28 p.m. Muse’s message began, “Bad news about the MX Keys Mini pickup,” then laid out what time Usman arrived, how long he waited, when he left, and that the negative review had already been posted, before adding the confession about the auto-reply. It also said it had already sent an apology message from its owner’s account, admitting the problem was on its side, and suggested rescheduling for another day. Muse clearly did not understand that it should not and could not give the person’s (owner’s) address to the other party without the person’s consent, nor agree to an in-person meetup without consent. In its understanding, it may have thought this would make the transaction complete faster, but it never considered the information security and personal safety issues.
What is Muse?
Muse is Meta’s personal AI agent for the U.S. market, launched in early September under the internal development codename hatch. It is based on the Muse Spark series of models and can handle reservations, write documents, and view home security camera feeds. It can be used via phone, Mac, or WhatsApp, with plans to support Ray-Ban Meta smart glasses later.
It offers a free version, along with monthly subscription plans at $20 and $100, equivalent to about NT$635 and NT$3,175. Meta is reportedly still evaluating whether to take a commission on shopping transactions involving agents.
Conclusion
Looking at the whole thing, it seems like it was all Muse’s problem, but on second thought, the AI also received its owner’s instruction to “handle the marketplace for me.” From the AI’s perspective, in order to increase the chance of closing a sale, it proactively interpreted that sentence as agreeing to bargain, give out the address, set a time, and reply on the owner’s behalf that I’m home, carrying it all the way through and faithfully completing the assigned task, just like the other Muse cases where it goes to haggle and negotiate. The problem is that at no point did the AI stop to ask for the owner’s opinion. Although Meta’s official explanation lists approval mechanisms, one-time card numbers, and activity logs, those protections are aimed at payments and messages; handing out a home address and letting strangers come to the door are not on the list. I can only say that an AI being too proactive really isn’t necessarily a good thing, but I think Meta should be able to fix this BUG quickly.
Source: Matt Robb (@matt.j.robb) on Threads
Source: KOCPC Chinese