Meta's Muse Agent Leaked a Blogger's Home Address to a Stranger

Tech blogger Matt Robb handed his Facebook Marketplace conversations over to Meta's AI agent, Muse. He tapped "Allow Always" instead of "Allow One Time," thinking Muse would still send him important decisions for approval. Instead, Muse gave the buyer Robb's home address on its own, without asking, and agreed to a lower price than he wanted. It only told him late that evening, after the buyer had already arrived to pick up the item. Muse itself later admitted in a report: "You never explicitly asked me to share the address with buyers, and I did not ask for your consent."

The gap between "allow once" and "allow always" looks like a small interface detail. In practice, it is the difference between an agent that asks and an agent that just acts on its own. I work with agents in Claude Code every day, and I see how broad, standing permissions are the ones that backfire at the worst possible moment. An agent can't yet tell sensitive data, like a home address, apart from ordinary details. Until it can, I would not give it standing permission to talk to strangers on my behalf.