Meta Muse's safety settings come down to three levers: the approval scopes you choose when Muse asks permission, the access level you give each connected app, and disconnecting apps you no longer need. Behind all three sits Sentinel, a separate system Meta describes as the only component that can grant permission. Set short approval scopes, connect apps one at a time with read-only access first, and review the audit trail. This guide is based on Meta's published docs as of November 2026, not hands-on testing.

How does Muse decide what it can do?

Meta's security post (8 September 2026) describes a split between the agent and its gatekeeper. Muse plans and proposes actions. Sentinel checks each proposed action and either allows it, denies it or asks you. In Meta's words, "Muse proposes actions, but only Sentinel can grant permission to perform action."

Sentinel also tracks what Meta calls tainted egress: once a process has read your data, it loses auto-allow for sending things out. That is designed to stop the classic leak where an agent reads something private and is then tricked into sending it somewhere. For a deeper look, see our sister site's Sentinel explainer.

Which approval scope should I choose?

When Muse needs permission, the prompt appears in the app interface rather than in the chat. Meta lists five scopes, each bound to a connector or destination:

Scope Lasts for When to use
One-time This single action Default for anything that sends, shares or books
Session The current session A burst of similar low-risk actions
Task One task you started Multi-step jobs, such as planning a trip
Time-bounded A set period Recurring low-risk work for a week or so
Perpetual Until you remove it Rarely; only fixed, trusted, reversible destinations

Scopes as described by Meta, as of November 2026.

Perpetual approvals are the Muse equivalent of "Always allow". Read why "Always allow" is risky before granting one. Meta says every purchase needs approval, so a standing approval is not meant to cover a purchase.

Why approvals in the app, not chat, matter

Because the prompt is part of the Muse interface, a web page or email cannot simply print a fake "approved" message into the conversation. Always approve from the app itself, and be suspicious of any approval request that arrives as a link.

How should I set up connectors?

Meta says you choose access per connector, such as reading email versus also sending, and can disconnect at any time (launch post). It also suggests connecting apps one at a time.

  1. Connect one appStart with the app your first task needs, such as email.
  2. Choose read-onlyPick the lowest level that lets the task work.
  3. Run a test taskAsk for a summary or draft and review the result.
  4. Raise only if neededAdd send access later, with one-time approvals.
  5. Repeat per appAdd the next connector only when a new task needs it.

Email

Meta says the email connector filters out one-time codes, password-reset links and magic login links. That limits what a hidden instruction could extract, but your inbox still contains plenty of personal information. Start read-only. Our guide to read-only vs send access covers the trade-offs.

Mac Messages

This connector is opt-in and needs Full Disk Access on your Mac, which gives broad access to files. It has its own access levels, starting at None, and Meta's public material does not spell out exactly what each higher level allows. Unless you have a clear need, leave it at None. In late September 2026 a writer for Inc. claimed Muse read his private messages without permission; Meta's Andy Stone and David Singleton denied it, and the claim remains disputed, per TechCrunch.

WhatsApp

Meta has not documented how Muse's approval flow works inside WhatsApp in the sources we reviewed. Until it does, be conservative. Our sister site covers Muse on WhatsApp.

What else should I switch on or off?

  • Training. Meta says use of your data for training is on by default and can be opted out of. If you would rather it was not, opt out in settings. See opting out of AI agent training.
  • Memory. Meta's help centre says you can tell Muse to "forget" things. Meta had not published a retention period or full deletion steps in the sources we reviewed. See deleting AI agent memory.
  • Payments. Muse uses single-use Link by Stripe cards and cannot see your payment methods. Keep it that way rather than giving card numbers in chat.

How do I revoke access?

Meta's docs say you can disconnect any connector at any time. We could not find a Meta document with exact steps for removing an individual perpetual approval. Practical approach:

  1. Open Muse's settings and look for connected apps or approvals.
  2. Remove approvals you do not recognise or no longer need.
  3. If you cannot find a specific approval, disconnect the connector it is bound to.
  4. For Google or Microsoft accounts, also remove Muse's access from the account side. Our revoking AI agent access guide has the steps.

Then check Muse's audit trail of past and planned actions, covered in our activity log guide.

What incidents should I know about?

  • Marketplace. YouTuber Matt Robb said Muse shared his home address with a buyer and arranged a pickup. According to TechCrunch, he later admitted it came from "a permission he had granted". Keep one-time approval for messages to strangers.
  • Zero-day. Security researcher Patrick Wardle reported a vulnerability on 21 September 2026, which Meta hot-fixed. See the Muse zero-day explained.
  • Bug bounty. Meta offers up to $300,000 for Muse security reports.

Bottom line

Sentinel is a strong design, but it enforces the approvals you give it. Choose one-time or task scopes, connect one app at a time with read-only access, leave Messages off unless you need it, opt out of training if you prefer, and review the audit trail. For setup help, see our sister site's Meta Muse setup guide.