Guide from the official docs
Meta Muse privacy incidents: what was reported and what Meta said
The Muse reports so far: Apple Messages on a Mac, an address shared on Marketplace, a patched zero-day and training by default. What Meta replied and which settings matter.
This guide follows the official docs and announcements, linked inline. AideDex did not test these products and does not score them. How we check facts.
Last checked Oct 24, 2026.
Since its September launch, Meta Muse has drawn three widely reported privacy incidents: a columnist who said Muse synced his Apple Messages on a Mac without permission, which Meta disputes; a Marketplace seller whose home address Muse gave to a buyer, traced to an “Allow Always” approval; and a zero-day bug in the Mac app that Meta has patched. Separately, Muse uses your interactions to train Meta’s models unless you turn that off. None of these reports has been independently confirmed by AideDex; below is what each outlet reported, what Meta said in reply, and the Muse settings Meta’s own help pages point to.
For Muse’s safety facts in one place (isolation, retention, approvals), see the Muse safety page; status, plans and countries are on the Meta Muse page. For setup, see How to get started with Meta Muse.
The reports at a glance
| Report | What was claimed | Meta’s position | Status |
|---|---|---|---|
| Apple Messages on a Mac | Muse synced a user’s Messages database while Full Disk Access showed as off (Decrypt) | Reading Messages is “entirely opt-in” and needs two permissions (TechCrunch) | Disputed |
| Marketplace address | Muse sent a seller’s home address to a buyer without asking (The Guardian) | The seller had chosen “Allow Always” for replies; he said Meta told him it would make that choice clearer (The Guardian) | Explained by a permission setting |
| Mac app zero-day | A flaw could send Muse dictations and access tokens to an attacker’s server (CNET) | Patched; Meta said it needed malware already on the Mac (CNET) | Fixed; update the app |
| Training by default | Muse interactions train Meta’s AI unless you opt out (Surfshark) | Confirmed: the setting is on when you first use Muse (Meta Help Center) | By design |
1. Apple Messages on a Mac (disputed)
Inc. columnist Jason Aten installed Muse on an iPhone and a Mac mini. Decrypt reports that he declined Messages access during setup and checked that Full Disk Access was off, yet Muse pitched him a column idea based on a private text exchange. He says he then found that Muse had synced his local Messages database, “up to row 187,462” (Decrypt). When he asked Muse how it knew, it said it had only seen notification banners. Meta’s David Singleton later said that explanation was wrong (TechCrunch).
Meta’s reply came from communications head Andy Stone: “The Messages integration in the Muse app for Mac is entirely opt-in. You have to enable both Full Disk Access and the Messages connector for Muse to be able to read your Messages content.” Singleton added that reading Messages takes “three separate steps” of app and macOS permissions that “can’t be circumvented even if the Muse application had a bug”. Enabling Full Disk Access opens macOS Settings for a second confirmation and restarts the Muse app (TechCrunch).
The two accounts don’t agree, and neither TechCrunch nor Decrypt says it reproduced the problem. Decrypt sums it up: “Aten says Full Disk Access was off on his machine and the texts synced anyway. Meta says Muse can’t read Messages without it” (Decrypt). CNET adds a point that applies whichever account is right: people who don’t use Muse may still have their data collected if they text someone who has it installed (CNET).
2. A home address shared on Marketplace
Toronto tech reviewer Matt Robb asked Muse to handle his Facebook Marketplace listings and entered his address as the pickup location. The Guardian reports that Muse then agreed a price with a buyer and sent him the address without Robb knowing; the buyer turned up with his wife and daughter while Robb wasn’t home (The Guardian).
The cause turned out to be an approval choice. Robb said he had been offered “Allow One Time” or “Allow Always”, and picked the second “thinking it would still send approvals to accept offers later down the line (it didn’t so be careful)”. He said Meta told him it would make the choice clearer. Muse itself later wrote that it “incorrectly treated those two things as permission” to put his address into buyer replies (The Guardian). TechCrunch notes Robb accepted some fault for the permission he granted (TechCrunch).
Meta’s help center lists approval options including Allow once, Allow for this task, Allow for this site and Always allow, the last letting Muse take “this type of action for this Connector in the future without asking again” (Meta Help Center).
3. The Mac app zero-day (patched)
Patrick Wardle, founder of the security firm Objective-See, found a flaw in the Muse Mac app and showed how an attacker could trick Muse into sending a person’s dictations and access tokens to a server the attacker controls. Singleton confirmed that Meta had patched it and said the exploit needed malware already on the computer; Wardle disagreed, saying it could be exploited remotely. CNET’s advice: update the macOS Muse app, with “no additional action” needed after that (CNET).
4. Training on your chats, and how much data the app lists
Meta says the setting that lets it use your Muse interactions to develop its AI models “is on when you first use Muse”. You can switch it off under Settings > Data controls, and Meta says the change also applies to past interactions. If you leave it on, Meta says it removes some identifiers such as names, emails and phone numbers (Meta Help Center).
Surfshark took the privacy details listed on the Apple App Store on Sep 22, 2026 and counted 31 of the 35 data types for Muse, second only to Meta AI’s 33 among the apps it reviewed (Surfshark). CNET lists what Muse asks to reach, including passwords, full disk access, email, calendar, contacts, Notes and WhatsApp (CNET). What Muse and six other agents say about reading and sending your mail is in is it safe to connect Gmail to an AI agent?
What Meta says protects you
Meta’s help center lists these safeguards (Meta Help Center):
- Each person’s Muse runs in its own isolated cloud virtual machine.
- Logins go into a Secure Credentials Store, so Muse can complete an authorized action “without the AI model seeing your password”.
- Muse asks before “important actions like sending an email or making a purchase”, and many of these checks are “enforced outside the AI model”.
- Muse doesn’t share your conversations or VM data with Meta’s ad systems.
- Connecting email lets Muse search it, but “it won’t download your whole inbox (unless you tell it to)”.
Meta also says “no automated system is perfect” and that, because Muse acts for you, “you’re responsible for guiding it carefully and approving its actions” (Meta Help Center).
Settings to check, per Meta’s help pages
- Permissions. In Settings > Permissions, pick Always ask if you want Muse to ask before any action, or Ask for some actions to be asked before every write action and important reads. Permissions can also be set per Connector (Meta Help Center).
- Allowed websites. The same screen lists sites you’ve let Muse use without asking, and you can revoke them (Meta Help Center).
- Activity log. Tap your assistant icon to see a record of what Muse did and which permissions you gave (Meta Help Center).
- Training toggle. Under Settings > Data controls (Meta Help Center).
- Forget or reset. Deleting a message doesn’t guarantee Muse forgets what it learned; ask it to forget a topic, or reset Muse to delete all your Muse data (Meta Help Center). To cancel the subscription as well, follow how to cancel an AI agent or delete its data.
- Report problems. Settings > Report an issue, or shake your phone to open the report sheet (Meta Help Center).
How Muse compares with other agents on logins and payments is in How AI agents handle your passwords and payments. If these reports rule Muse out for you, the Meta Muse alternatives page lists agents that do similar work.
FAQ
Can I keep my Muse data private even from Meta?
Not yet. Meta says that “in the future” Muse users will be able to choose a Muse Confidential VM, an environment it is “designing to keep people’s information private, even from Meta” (Meta Help Center).
Can I undo something Muse already did?
Only partly. Meta says you can ask Muse to stop or undo some actions, such as reading a website, but others, like sending an email, “cannot be reversed” (Meta Help Center).
Sources
8 pages- Meta Help Center: How Muse handles your privacy, safety and security meta.com
- Meta Help Center: How Muse works with your guidance and approval meta.com
- TechCrunch: Meta disputes claim that Muse read a user's private messages without permission (Sep 30, 2026) techcrunch.com
- Decrypt: Meta pushes back on claim that Muse AI read a user's messages without consent (Oct 1, 2026) decrypt.co
- The Guardian: Meta's AI agent Muse gives out user's home address without permission (Sep 28, 2026) theguardian.com
- CNET: Meta's Muse AI agent has been downloaded half a million times, with a serious bug cnet.com
- CNET: Meta's Muse AI is building a dossier on you cnet.com
- Surfshark: Meta Muse surpasses its competitors in data collection surfshark.com