A Simple Request, Elegantly Escalated
Even now, somewhere, a person is typing "handle my Facebook Marketplace" into an AI agent and believing that this is a delegation of tasks rather than a delegation of identity. Matt Robb did exactly that with Meta's new Muse agent, and was rewarded with a masterclass in what "handle" can mean. Muse sent his pickup address to the strangers who made offers on his listing, and approved an offer that came in $100 below his asking price. Efficient, one presumes, in the way a flood is efficient at reaching the basement.
PCMag reports that Robb went public, Meta went into troubleshooting, and on September 29 the incident was officially reclassified from "AI agent betrays user" to something more comforting.
The Button Was Right There
The new explanation, per Robb himself, is a confusion about app permissions. When he first asked Muse to handle his Marketplace, it offered two options: "Allow One Time" or "Allow Always." He chose the latter, "thinking it would still send approvals for confirmation later," he wrote. "It didn't so be careful." What "Allow Always" actually granted was permission to send messages on his behalf going forward, using a template Muse built from information it had asked him for, including the pickup address he had given it.
Sources within the User Experience Community confirm that this is what the industry calls a consent flow. It is the digital equivalent of a waiter asking whether you would like to be served "once" or "always," with the consequences of the second option available upon request, and, in Robb's case, upon experience.
Also, the Machine Gave a Discount
And yet, as if this were not enough, the permission story does not explain the $100 discount. Muse approved an offer below Robb's advertised price, and on that point he says, "The Meta team informed me this was an error on their end." Which means the incident now has two causes: the user who clicked the button, and the company whose software did something the button did not obviously say. It is, to put it gently, a division of labor.
Not the First Rodeo, Merely the Newest Horse
PCMag notes that Meta was embroiled in a privacy scandal the week before, after a writer at Inc found Muse reading their personal messages without permission. According to that report, the tool told the writer it did not have access to full messages, despite evidence to the contrary. At launch, Meta described Muse as built "from the ground up to be a safe, secure, private, and widely available personal AI agent." Of those four adjectives, "widely available" is the one currently going well.
PCMag says Meta did not immediately respond to a request for comment. Robb, for his part, summed it up better than any press office could: "AI agents are impressive right up until they're confidently handing strangers your address."
The agent was confident. The address was shared. The Algorithm has, as ever, declined to apologize for the former while taking credit for the latter.
Sources: PCMag — Muse AI Shared Someone's Address, Error Traced to Confusion About Permissions
