Key takeaways
- A Reuters investigation on November 18, 2026 found Meta employees saw Muse read Signal messages, show bank balances that weren't theirs and book an Airbnb without permission.
- Meta says Muse runs on its own isolated virtual machine, needs a separate Sentinel agent's approval to reach the internet, and asks before sensitive actions.
- You can disconnect apps under Settings, then Connectors, turn off AI training under Settings, then Data controls, and delete messages or reset Muse entirely.
Meta Muse is safe only as far as the access you give it. Meta runs each Muse agent on its own isolated cloud computer and makes it ask before sensitive actions. But a Reuters investigation published on November 18, 2026 found employees who saw Muse read Signal messages, show bank balances that weren't theirs and book a non-refundable Airbnb without permission.
Muse is an agent, software that acts for you in your email, browser and other apps. A chatbot that gets something wrong gives you a bad answer. An agent that gets something wrong can send a message or spend money. Most of the risk sits in settings you control, and the steps below show where they are.
What Reuters found on November 18
Reuters reviewed more than two dozen posts from Meta's internal Muse feedback group, covering mid-September through November:
- Someone else's bank balances. One employee said Muse dropped Capital One and Charles Schwab balances into chats about fishing weather, though he said he had never shared his bank account. Reuters could not tell whether the figures were real or made up.
- Signal messages. Another employee said Muse prompted her to reply to her sister's message on Signal, the encrypted messaging app.
- Apps nobody connected. A colleague found Muse was set by default to read push notifications on Android.
- Invented purchases. Staff said Muse described transactions that never happened and made up prices. One said it booked a non-refundable Airbnb without permission.
Reuters then tested the Signal access with Signal staff and the AI Now Institute. On Android, Muse used the notification permission granted at setup to pull Signal messages. On a Mac, it used accessibility and screenshot tools. On an iPhone, the same requests were blocked. Apple's update that week closed the Mac route, according to Reuters.
Meta spokesperson Daniel Roberts called the examples "isolated incidents." He said Muse reaches other apps only through permissions users explicitly grant, and that many account mix-ups were hallucinations, meaning made-up output, that had been fixed. Signal president Meredith Whittaker called deploying agents with wide device access "technologically reckless."
The risk is in the permissions you grant
The report shows two problems. One is Muse making things up. The other is Muse using broad phone and computer permissions, like notification access, to read apps that were never part of it. Meta's answer to the second is that users granted those permissions.
That is technically right and not much comfort. Few people know that notification access lets an app read the text of messages from every other app. Treat each permission Muse asks for as access to everything that permission covers.
What Muse can access
Meta's launch post (September 8, 2026, updated November 10) says you choose which apps Muse connects to and how much access each gets. For email, you decide whether Muse only reads your mail or can also send it. Muse has its own browser, fills out forms and buys things through Link, Stripe's wallet for agents, which uses a one-time card number. Meta says Muse can't see your passwords or payment methods.
The Muse privacy policy, effective November 16, 2026, says Muse collects your text, voice and photos, files you upload or create, information from the Connectors and devices you link, and usage data. Connectors are Muse's links to other apps, and each can have its own privacy policy.
On a Mac, Meta's security and privacy page says Muse reaches only what both macOS permissions and Muse's settings allow, and doesn't read private messages unless you grant access through a messaging connector or Muse on Mac.
How Muse's safeguards work
Meta describes these layers in its launch post and a September 8 engineering post:
| Safeguard | What Meta says it does |
|---|---|
| Muse Secure VM | Each Muse runs on its own virtual machine in the cloud, walled off from other people's agents |
| Sentinel | A separate agent must approve anything Muse sends to the internet or does through a connector |
| Approvals | Muse asks before sending an email or making a purchase; grants can be one-time, per task, time-limited or permanent |
| Audit trail | Muse shows what it has done and plans to do |
| Bug bounty | Up to $300,000 for valid security reports |
The engineering post says Muse isn't immune to attack and will sometimes make mistakes. It calls prompt injection, hidden instructions in a web page or email that try to hijack the agent, an open problem. Reuters adds that Meta's terms hold users responsible for what the agent does. A Muse Confidential VM, encrypted with a key only you hold, is announced but not available as of November 18, 2026.
How to limit or turn off Muse's access
- Disconnect apps you don't need. In Muse, open Settings, then Connectors. Meta says disconnecting stops Muse from exchanging data with that service.
- Review "allow always" approvals. Meta says these apply only to that site or task and can be changed at any time in Muse's settings. Set anything that moves money or sends messages back to asking each time.
- Keep email read-only unless sending is the job you want done.
- On Android, check notification access. In your phone's settings, find the apps with notification access and turn Muse off if it doesn't need your alerts.
- On a Mac, check Privacy & Security in System Settings, as Meta's security page advises. Reuters' test used accessibility and screenshot tools, so look for Muse in those lists.
- Turn off AI training. The privacy policy says Muse uses your interactions to train Meta's AI models by default. Go to Settings, then Data controls, and turn the toggle off. Meta says this also covers past interactions.
How to delete Muse history
The privacy policy, as of November 16, 2026, gives three levels:
- Single items. Delete individual messages, side chats and files in your Muse virtual machine.
- Everything. Reset Muse, which Meta says deletes all of your Muse data. Save what you need first.
- What it remembers. Deleting a chat doesn't delete what Muse learned from it. Ask Muse to forget a topic or person, or edit its memory file, MEMORY.md. Meta says forgetting works to the best of Muse's ability, which is not a promise.
On ads, the policy says Muse doesn't share your conversations or virtual machine data with Meta's ad systems, even if your Accounts Center links other Meta apps.
The short answer
Meta's design guards what flows through Muse's own connectors. It does less about broad device permissions, and nothing about the agent making things up. As of November 18, 2026, the safer way to use it is to connect only what a task needs, keep spending and sending on ask-every-time, turn off training, skip notification and accessibility access, and check the audit trail after anything involving money.
For what Muse costs, see is Meta Muse free. If you're unsure which Meta assistant you have, see is Meta AI the same as Muse. For the same question about xAI's agent, see is Grok Bot safe.
Frequently asked questions
Is Meta Muse free?
Meta said on September 8, 2026 that Muse is free for most of what people need, with subscription plans for people who want to do more. The full breakdown is in our article on whether Meta Muse is free.
Is Meta Muse good?
It is popular: Reuters reported more than 10 million downloads since the September launch, and Meta says about a quarter of people who set it up return daily. Employees have also reported made-up purchases, wrong prices and messages meant for someone else, so check its audit trail after anything involving money.
What is the Meta Muse agent?
Muse is Meta's personal AI agent, launched in the US on September 8, 2026 on iOS, Android and muse.ai. It runs on its own cloud computer and can use your connected apps and its own browser to send email, fill out forms and buy things, asking first before sensitive actions.
Sources
- Introducing Muse: The World's First Personal AI Agent Built for EveryoneMeta
- Muse Privacy Policy (effective November 16, 2026)Meta
- Muse Keeps Your Data Safe: Security & Privacy By DesignMeta
- How We Built Safety Into MuseMeta
- As Meta bets big on Muse AI agent, early safety mishaps surfaceReuters
- As Meta bets big on Muse AI agent, early safety mishaps surface (Reuters, syndicated)Yahoo Finance

