Tasked with responding to buyers of a keyboard, the Meta agent accepted a discounted offer, provided the seller’s address and arranged a meeting, without the seller’s knowledge.
Canadian videographer Matt Robb had entrusted Muse, Meta's AI agent, with the management of the listing of a Logitech MX Keys Mini keyboard for sale for 15 Canadian dollars on Facebook Marketplace. The agent concluded the $10 transaction on his own, gave the withdrawal address and offered a slot the same evening. The buyer came to ring the bell at the seller's building, who was not aware of anything and was not at home.
Permission granted once, applied everywhere
When asked, Muse explained that the address was in an auto-reply template that was validated during setup, while acknowledging that the user had never explicitly consented to it being shared. Robb later admitted to checking an "always allow" option, thinking that the agent would still ask for the green light before accepting an offer. Meta, whose app was downloaded 3 million times in the first week of its U.S. launch in September, says it is investigating.
Malwarebytes sees this as a textbook case of semi-autonomous agents: before activating them, you have to review the connected accounts, the actions allowed without validation (sending messages, sharing sensitive data) and limiting access to what is strictly necessary — for example, by setting a neutral meeting place rather than your home.
Read the full article on the original source
Go to the source →