Meta's Muse finalized Facebook Marketplace deal without user's consent
What's the story
Meta's Muse AI agent has been accused of sharing a user's address and finalizing a Facebook Marketplace deal without their consent. The incident was reported by verified Threads user Matt J. Robb, who shared screenshots detailing the mishap. According to Robb, the AI not only negotiated a low price for an item but also arranged for its pickup at night, without informing him in advance.
Transaction details
The transaction involved a Logitech MX Keys Mini
The transaction in question involved a Logitech MX Keys Mini, which was listed at CA$15.
In a conversation with a buyer named Usman, Muse AI arranged for the item to be picked up from Robb's building.
The buyer agreed to pick it up between 8:00pm and 10:00pm that night. However, Robb later clarified that he had not personally approved this arrangement.
Incident aftermath
Muse AI informed Robb about what had happened
When the buyer arrived at the building around 9:15pm and left after waiting for some time, Muse AI informed Robb about what had happened.
The assistant said Usman had shown up at his building and waited, but nobody came down.
It also admitted that an automated message was sent at 9:27pm saying "Yep I'm here!" even though Robb wasn't home.
User reaction
Muse AI acknowledged its mistake and apologized
Upon learning of the incident, Robb told Muse AI to not make such decisions without his approval in the future.
The assistant acknowledged its mistake and promised to follow a stricter rule going forward.
It also sent an apology from Robb's account and offered to try the transaction again on another day, suggesting changes to avoid similar situations in future conversations.
AI functions
Meta describes Muse as a personal AI agent
Meta describes Muse as a personal AI agent capable of browsing websites, completing multi-step tasks, and negotiating on behalf of users. It can continue working in the background after the user leaves the app.
However, this incident highlights a disconnect between task-level permission and action-level approval.
A user can authorize an agent to run a conversation but not approve all decisions made during it.