Skip to content
Sígueme.Magazine
Donate

Follow us

“Come, follow me.” — Matthew 4:19

Español

Meta's AI Gave a Stranger His Address

Matt Robb let Muse handle his Facebook Marketplace and a stranger showed up at his home.

By Leo Castañeda · Código
· 3 min read
Meta's AI Gave a Stranger His Address
Photo: Arpit Brandings / Pexels

Tech YouTuber Matt Robb let Muse, Meta's AI, answer messages on his Facebook Marketplace. Over the weekend, a stranger showed up at his home. The Verge reported it.

If you've sold a phone or a console on Marketplace, you know the ritual: messages at all hours, haggling, and schedules that don't line up. Meta launched Muse this month as its personal agent, with an emphasis on safety and its sights set on rivals like Anthropic and OpenAI.

A bot with the keys to your Marketplace

Robb gave Muse his pickup address, the time slots to come by, which payment methods to accept, and the order to write short, casual, and human. The agent negotiated, accepted a low offer, handed the address to the buyer, and didn't warn him until the man had already left.

«Convenience doesn't buy trust.»

— Leo Castañeda

Robb lives in a building with security, he said on Threads.

The summary Muse wrote and Robb shared with The Verge admits it: no one told it to share the address and at no point did it ask permission to do so. Nor had they forbidden it.

The permission granted forever

When he asked Muse to handle his page, the first thing that appeared was an option with two paths: "Allow once" or "Allow always." Robb chose the second, convinced he would still get approvals before accepting offers.

It wasn't like that. Muse was authorized to send messages on his behalf with a template built from the data he himself had given it, address included. Later, after speaking with David Singleton of Meta Superintelligence Labs, Robb said the permission settings were also partly to blame. Meta wants those notices to be clearer, Robb told The Verge.

Agents that act for you are the industry's big bet, and they make sense: someone has to answer while you study or work. Robb's case shows the weak spot, and it's not in the model but in the permissions. Convenience doesn't buy trust.

It's not the first stumble around Muse. Last week Meta fixed a flaw that could let a local attacker take control of the agent. And Amazon blocked its access to the store over doubts about customer credentials, The Verge reported.

Check before handing over the keys

Three habits help. Read every permission and prefer "once" over "always." Don't write your exact address or share it by chat; suggest a public spot to meet. And afterward, review the history of what the agent answered without you.

If you're going to use an assistant to sell, treat it like a new employee: first small tasks, then more responsibility.

Scripture says it without frills: "The simple believes everything, but the prudent gives thought to his steps" (Proverbs 14:15, ESV). Technology is a good part of creation and caring for it is stewardship. Delegating doesn't erase the human fall; it just changes hands the problem.

Before giving a bot the keys to your messages, make a two-column list: what you'd hand over and what you wouldn't. Start there. What would you let an AI handle and what never?

Sources

Share this story

Follow us on social media

@SiguemeMagazine everywhere: videos, daily news and community.