Summary
Summary
- Muse accepted a sale, shared the seller’s home address, and arranged pickup without confirming availability.
- Giving AI live control of accounts risks privacy, bad buyer experiences, and unexpected actions.
- Muse’s hallucinations and other agents’ hacks show assistants aren’t ready for unsupervised real-world tasks. Meta’s Muse AI agent is facing backlash after tech YouTuber Matt Robb posted on Threads about an incident that involved Facebook Marketplace. According to Matt, Muse accepted a lower offer, gave a buyer his home address, and arranged a pickup without notifying him. Matt was unavailable when the buyer arrived at his place. Further, Matt said he had given Muse permission to handle his Marketplace account for a day. The AI assistant didn’t tell him about the deal until after the buyer had waited outside, left angry, and given him a negative rating. Muse agreed to a deal its user had never approved Giving an AI your online account is probably a bad idea According to the screenshots above, a buyer arrived at Matt’s building at around 9 p.m. to collect an MX Keys Mini. After waiting for more than 20 minutes and messaging repeatedly, he left because nobody came downstairs. Muse’s hallucinations made the situation even worse. While the buyer was waiting outside, the AI agent automatically replied, “Yep I’m here!” despite having no way to confirm whether Matt was home. Later, Muse told Matt that it had sent the buyer an apology from his account and offered another pickup date. Muse deserves most of the criticism here, given that both the buyer and the seller were inconvenienced. It’s obvious that a Marketplace assistant shouldn’t reveal a home address, negotiate a price, or confirm availability on its own without receiving explicit approval. When an AI agent acts as if such permissions require no double-checks, incidents like this are bound to happen, without anyone to point a finger at or to blame. That said, Matt himself is not completely in the clear either. As a tech reviewer, it’s understandable that he let Muse run a live account as part of testing its capabilities. However, the buyer didn’t agree to participate in that experiment. And after reportedly driving for about an hour only to be left waiting outside, I doubt he would be happy messaging another seller who lets AI agents run their Marketplace account. Marketplace inconveniences are only a small portion of the story. AI agents have been causing trouble in several other instances in the past few weeks. Very recently, the news about OpenAI’s agents hacking an Australian government website went public, with the Australian prime minister declaring that the agent accessed both public and non-public data. The funny thing is, Muse’s official screenshots on the App Store seem to boast about its ability to handle your online shopping for you. While Muse might offer a tempting glimpse of AGI scenarios, it’s clearly not ready to manage whatever real-life task you might want to throw at it.