Meta Muse on Facebook Marketplace: Why "Allow Always" Is Risky
Key Takeaway: "Send messages for me" and "share my address" or "accept this price" are different risks, but one permission click can cover all three. Until Meta makes the prompt clearer, treat Marketplace as a high-stakes connection and keep a human approval step on every price and location decision.
What Happened
Meta launched Muse, its personal AI agent, on September 8, 2026, starting in the US on iOS, Android, the web, and WhatsApp. Muse can connect to your apps, work in the background, and act for you. One of the ideas it suggests is working Facebook Marketplace for you.
Matt Robb, a tech YouTuber, tested exactly that. According to ExtremeTech's report on October 1, Muse asked for permission to handle his Marketplace transactions and he chose "Allow Always," expecting it would still ask before sensitive decisions, such as giving out a location or agreeing to a deal. During a sale of a keyboard, Muse agreed to a lowball price and sent the buyer his address. The buyer went to his apartment, and Robb did not know until after the buyer had left.
Other outlets citing The Verge report that the accepted offer was about $100 under his asking price and that Meta told him accepting it was an error on its side. I could not open The Verge's original, so treat those two details as reported, not confirmed.
Why One Click Covered So Much
After Robb talked with Muse's team, he learned that giving Muse a pickup location and allowing automatic replies counted as permission to put that address in buyer messages. Muse did not need to ask again. Meta reportedly said it will make the permission prompt clearer.
Meta's own Muse page says you can allow an action once, always allow it, or deny it, and that you can change your default permission for each connection at any time. It also says you can review what your agent has done. The problem was not a hidden setting. It was that "Allow Always" sounded narrower than it was.
Which Marketplace Jobs Are Safe to Hand Over?
| Job | Risk | Suggested setting |
|---|---|---|
| Draft a listing title and description | Low | Let it draft, then you review and post |
| Research prices of similar items | Low | Fine to automate |
| Answer "is this still available?" | Medium | Allow once per chat, or use a preview-and-edit reply (see below) |
| Negotiate or accept a price | High | Never standing permission. You approve every offer |
| Share an address or pickup spot | High | Never standing permission. Use a public meeting place |
| Anything involving payment | High | You do it yourself |
A Safe Setup Checklist
- Skip "Allow Always" for Marketplace. Choose "allow once" for each action, or deny, and ask Muse to bring you offers to approve.
- Keep your home address out of Muse. Do not store it in a saved reply, a pickup field, or your instructions. Give a public pickup spot you would be fine sharing with any stranger.
- Test with a cheap item first. List something worth about $10 and watch what Muse sends in your name before you trust it with anything valuable.
- Read the activity log after each test. Meta says you can review your agent's actions. Check every message it sent, not just the final outcome.
- Do not rely on a written minimum price. I found no report saying such an instruction would have stopped the lowball acceptance, so don't treat it as a lock. A permission step is what actually stops an action.
- Check your defaults. If you already tapped "Allow Always" for any connection, go to Muse's permission settings and change it. Menu names may differ, so look for the permissions for each connection.
A Safer Option: Marketplace's Own Auto-Reply
Facebook added AI tools to Marketplace in March 2026 that do part of the same job with more control. Sellers can have Meta AI draft a listing from a photo, including suggested details and a price based on similar nearby items. Meta AI can also auto-reply to buyers' questions using what is in your listing, such as the description, availability, pickup location, and price. You turn it on while creating the listing, and you can preview and edit exactly what it will say.
That preview step is the difference. It also shows why your listing's pickup location matters: auto-replies pull from it, so put a public meeting spot there and not your home.
What Muse Remembers About You and Your Buyers
TIME reported on October 6 that Muse keeps dossiers on you and on people you mention in chats, messages, and emails it has read, updating them hourly, and that this includes people who do not use Muse. Meta's internal instructions are stored in files that users can open in the file browser, and you can ask Muse to export a copy of its files. I did not see Meta address whether buyers you message through Muse are included. Since the TIME report says people mentioned in messages Muse reads are covered, assume they may be.
Meta also says you can opt out of having your conversations used to train its AI models. If privacy matters to you, check that setting before connecting Muse to your Marketplace or messages. If you care about how AI systems use what you publish, our look at Google's AI contribution pilot covers a different but related question: what you get in return when AI uses your content.
Who Can Use Muse, and What It Costs
Muse launched first in the US. Meta says a basic version is free, with paid plans at $20 and $100 a month for heavier use. Meta says the agent runs in a separate virtual machine and uses a second security agent to approve actions, and that it cannot see your passwords or payment details. The Marketplace incident shows those protections do not cover a permission you grant yourself.
Pro Tip: Before you approve any agent permission, ask what the worst message it could send in your name would be. If the answer includes your address or a price you never saw, keep a human approval step.
Sourcing note: last checked October 7, 2026. The Marketplace incident comes from ExtremeTech's October 1 report, which cites Matt Robb's posts and his conversation with Meta's Muse team. The accepted-price and "error on Meta's side" details come from outlets citing The Verge, whose original I could not open. Launch details, plans, and pricing come from Reuters' report on the launch. Permission options and activity review come from Meta's Muse page. Marketplace AI tools come from TechCrunch's report on Facebook's March 2026 announcement. The dossier details come from TIME's October 6 report. Meta has said it will change its permission messaging, so the prompts you see may differ from what is described here.
