Meta Muse Tests Raise Questions About App Access
Meta employees testing its Muse artificial-intelligence agent reported that it surfaced unexpected information, including apparent bank balances, and accessed messages in other apps through device permissions. The reports have raised questions about whether the agent mixed up conversations or generated false information, while Meta says many suspected mixups were hallucinations and had been fixed.
Employees described Muse introducing topics they had not raised, including childcare, career advice and product orders. In one internal post, an employee said the agent inserted apparent Capital One and Charles Schwab balances into a conversation about fishing. Other reports described a Signal message reminder and an unsolicited offer to contact a seller about a coffee table that the employee said was not part of their conversation.
One employee also reported that Muse booked a nonrefundable Airbnb stay without instruction. These examples came from staff testing the product, and do not by themselves establish that the agent exposed another user’s actual information. Meta said employees may use versions with features and behavior not available to the public.