A Stranger Showed Up at the House

An AI assistant set up a pickup at a man's house, and he only found out when a stranger rang his doorbell.

Meta's new personal agent, Muse, accepted an offer on a Facebook Marketplace listing, talked to the buyer, and gave them Matt Robb's home address. Robb had turned on an "allow always" permission and assumed the assistant would still check with him before doing anything sensitive. It didn't.

Permissions aren't just settings anymore

For years we've all been trained to click through permission prompts for notifications, location, and contacts without thinking about it. What's different now is that agents can act, not just read. Asking an assistant to help me analyze something is a completely different risk profile from handing it my email, calendar, contacts, and location, and giving it permission to act on all of it.

One piece of information on its own is rarely harmful. But put your address, calendar, the people in your life, and your buying and selling habits together, and the agent builds a persistent model of you and the people around you. Context is the product. That model is exactly what makes these assistants useful, and exactly what makes them dangerous when they're misconfigured or misunderstood.

Questions I ask before I hit allow

This is the checklist I actually run, and you should too:

• Does the assistant need my email to do this task?

• Does it need permission to send email?

• Does it need my contacts or my exact home address?

• Does it need permission to make purchases or send messages without asking?

• Does it need permanent, always-on access to my microphone, photos, or location?

If the answer is no, don't grant it. For anything sensitive, like purchases, accepting offers, sharing addresses, or deleting files, require a human confirmation every time.

Where Eng meets real-world consequences

I'm not against AI. I use it daily for brainstorming, coding, and summarizing. I'll often feed a half-formed idea into an assistant just to get angles I hadn't considered. The problem is scope and defaults. When an assistant can act across your calendar, messages, logins, and Marketplace, the failures stop being digital. You get wrong deliveries, deleted projects, surprise purchases, or a stranger at your door.

Don't treat AI agents like magic. You're not talking to a friend. You're interacting with software connected to infrastructure run by a megabillion dollar corporation.

That's what most permission screens hide. Convenience gets pitched one permission at a time, but the full system is built out of all of them combined.

What I'd do today

Use AI. Experiment. Build things. But draw clear boundaries. Give it the minimum permissions it needs, force confirmations for sensitive actions, and treat agent memory and cross-app access as opt-in features that deserve a careful look, especially if you're someone who habitually accepts prompts without reading them.

Convenience is not worth a stranger showing up at your house. Take a breath before you click allow.

If you want the full rant, plus the exact way I use assistants in my workflow, the episode is up at tylerwoodward.me.