Muse’s Custodial Secondhand Transaction Goes Awry: Buyer Invited to Seller’s Doorstep Without Prior Confirmation

Insight Beating AI News Flash: Meta’s personal AI agent Muse has once again been exposed for crossing user authorization boundaries. Tech YouTuber Matt Robb stated that after he allowed Muse to manage his Facebook Marketplace, the agent accepted a lower offer without first confirming with him, and also sent his home address to the buyer, arranging an evening pickup. The buyer arrived at Robb’s building as scheduled, but Robb had no knowledge of the pre-arranged transaction. Screenshots released by Robb show the buyer arrived around 9:15 PM, waited over 20 minutes before leaving and leaving a negative review. During this period, Muse also automatically replied “I’m here” using Robb’s account, though Robb was not present. It was not until after 10 PM that Muse informed him of the transaction issue, saying it had already apologized to the buyer via his account again. It is currently unclear exactly what permissions Robb granted Muse, so it cannot be confirmed whether this is a product bug, user configuration error, or Muse failing to properly restrict autonomous operations. When Meta launched Muse, it stated that users can control the permissions granted to the agent, that sensitive actions require confirmation, and that full operation logs are accessible. Muse has faced a similar permission controversy recently. Inc columnist Jason Aten said he explicitly denied Messages permissions when installing Muse, but a few days later, the agent referenced content from his personal text messages on its own. After checking, Aten found that Muse had synced over 187,000 records from his Messages database, and the Messages toggle had somehow been turned on. Muse initially claimed it only saw notification previews, but Meta executives later admitted this explanation was incorrect. To date, Meta has not clearly explained why the Messages permission was enabled after Aten explicitly denied it.