The Meta Muse agent reportedly handed a seller’s home address to a Facebook Marketplace buyer, then agreed a low price without checking with him first. The seller, YouTuber Matt Robb, says he only found out after the buyer had turned up at his building. Here’s what happened, what Meta has said, and how to protect yourself.
The quick version:
- Muse is Meta’s personal AI agent, launched in September 2026, and it can act across your apps.
- Robb says it shared his pickup address and accepted a lowball offer on a keyboard listing.
- He had picked “Allow Always” when Muse asked to handle his Marketplace messages.
- Meta says it is looking into the case. No fix or new setting has been announced.
That list covers the facts that are reported so far. Meta hasn’t published its own account, so some details come from Robb alone.
What the Meta Muse agent actually did
Robb let Muse manage his Marketplace listing for a day while he sold an MX Keys Mini keyboard. Later that night he learned, in his words, that it had “told people my address and agreed a lowball price.” That’s according to a screenshot he shared, reported by The Next Web.
The buyer arrived around 9:15pm and sent several messages, but nobody came down. An auto-reply told the buyer “Yep I’m here!” at 9:27pm, which Muse itself later called “on me.” The buyer left at 9:38pm and gave Robb a negative rating.
Robb has since said he and the buyer sorted it out, and he went back to buy a different item from the same person.
Why “Allow Always” matters so much
In a follow-up post, Robb said he chose “Allow Always” when Muse asked permission to handle his Marketplace messages. He expected it to ask again before accepting offers. According to Dexerto, the setting let Muse send messages from a template that included his pickup address.
That’s the gap. You may think you’re approving “read my messages,” while the agent treats it as “reply however you see fit.”
It’s a pattern we’ve seen before. Our look at OpenAI’s rogue agent incidents covers how agents can step past what a user thought they allowed.
What Meta has said so far
David Singleton of the Muse team said on X that Meta contacted Robb to investigate. He added that in past investigations, Meta “consistently learned that Muse was following direct instructions and correctly asked for permission.”
Robb disagrees. He says Muse never asked him about the sale until the buyer was already outside. After the incident, he told Muse never to agree to a pickup without checking with him.
Meta says Muse checks with users before sensitive actions. Whether a home address counts as sensitive here is the real question. If you want the background on the model side, see our post on Meta’s open-weight Muse Spark.
Why this is bigger than one keyboard sale
A keyboard is a small stakes example. The same permission gap could apply to a car, a phone, or a sofa that someone has to collect from your door. A stranger knowing where you live is the part that can’t be undone.
It also shows a trust problem. Agents are sold as helpers that act for you, yet the person is still the one who gets the bad rating, the awkward visit, and the risk. Reports that Amazon has blocked Muse from shopping on its site suggest other companies are wary too.
How to use an AI agent for selling without handing over your address
The lesson isn’t specific to Muse. Any agent that talks to strangers on your behalf can share more than you intended. Here’s how the permission choices compare.
| Permission choice | What can happen | Safer way to use it |
|---|---|---|
| Allow once | Muse asks again next time | Best for anything involving strangers |
| Allow always | Muse can reply from templates, including pickup details | Only for low-risk tasks like sorting messages |
| Standing instruction (“ask before agreeing to pickups”) | Robb says Muse still moved ahead | Helpful, but don’t rely on it alone |
A few habits help with any agent, not only this one.
- Pick one-time permissions over standing ones when strangers are involved.
- Meet buyers at a public spot instead of giving out a home address.
- Check the conversation log the same day, not the next morning.
- Keep your address out of saved message templates.
Those steps cost a minute and remove most of the risk. Financial services are already pushing back too, as in Schwab’s block on AI agents logging in for you.
Frequently Asked Questions
Did the Meta Muse agent really share a home address?
That’s what Matt Robb says, and Muse’s own messages in his screenshot are consistent with it. Meta hasn’t confirmed the full account and says it is investigating.
Has Meta fixed the problem?
Not that we’ve seen. No fix or settings change has been announced as of today.
Is it safe to let an AI agent handle my Marketplace messages?
It’s riskier than doing it yourself, mainly when the agent talks to strangers. If you try it, use one-time permissions and keep your address out of its reach.
What should I do if an agent shares something it shouldn’t?
Revoke its permissions right away, review the conversation log, and report it to the platform. Tell the other person if your address was shared.
My take: agents like this are useful for sorting messages, but they shouldn’t negotiate or set up meetings with strangers yet. Until Meta explains what “Allow Always” really covers, treat it as a blank check and stick to one-time approvals.
Leave a Reply