Meta's AI gave a stranger his home address

Meta’s new Muse AI agent handled a Facebook Marketplace sale for a guy named Matt Rob, and it gave a complete stranger his home address without asking first. The guy shows up expecting to buy a keyboard, and Rob has no idea any of this happened until there’s a knock on his door. That’s not a hypothetical risk, that’s a thing that actually occurred, and it happened because of one “allow always” permission he clicked and forgot about.

I get into why this story bothers me so much, and it’s not because I think AI is evil or useless. I use it, I build with it, it’s a tool like anything else. The problem is the gap between using AI to help you think and handing an AI agent the keys to your email, contacts, location, calendar, and the ability to talk to strangers and make decisions on your behalf. Those are two completely different risk profiles, and Meta is rolling out the second one to a user base that includes a lot of people who have spent twenty years clicking “allow” without reading a single word of what they agreed to.

I walk through what Muse actually does, what Meta says about it, why I don’t buy the reassurances given their track record, and the specific questions you should be asking before you give any AI agent permission to act as you. If you’ve ever connected an app without reading what you were agreeing to, which is all of us, this one’s worth your time.