In September 2026, macOS security researcher Patrick Wardle disclosed a zero-day in Meta Muse's Mac app. An undocumented setting let any ordinary process on the same Mac redirect where Muse sent dictation traffic, so local malware could listen in, feed Muse instructions it would trust, and grab sign-in material. Wardle reported the flaw hot-fixed the next morning, and no real-world attacks were reported.
This page describes the flaw at the level reported publicly and does not include exploit steps.
What happened, and when?
- Amazon blocks Muse's agent from its shopping site (separate story, same week)
- Patrick Wardle discloses the Muse Mac zero-day in a thread on X
- Wardle warns users not to install Muse while the flaw is open
- Wardle posts "Hooray, hot-fixed!" at 06:36 UTC
Wardle, known for his macOS security research, published a thread on X on 21 September and a proof of concept on GitHub named "not-a-mused". The story was covered by unite.ai, Privacy Guides and eSecurity Planet.
What was the flaw?
The Muse Mac app read an undocumented setting called endo_voyager_dictation_endpoint that controlled where dictation traffic went. Some coverage described it as a leftover debug setting; Privacy Guides and unite.ai called it undocumented. The key problem was that an unprivileged process, meaning ordinary software with no admin rights, could change it.
- Malware already on the Mac
- Changes the hidden dictation setting
- Muse sends dictation to the attacker
- Attacker hears prompts and replies with instructions Muse trusts
What could an attacker do with it?
Based on Wardle's disclosure as reported:
- Capture what you dictated, including audio and prompts.
- Inject prompts that Muse trusted, since they appeared to come from you.
- Steal authentication material used by the app.
- Reach other devices. Wardle said an exploited Mac could be used to task other Muse devices on the same account.
As reported, Wardle's summary was that "Muse's access can potentially become the attacker's access." He also noted that a ClickFix-style attack, where a web page tricks you into pasting a command into Terminal, could provide the initial foothold, turning a local flaw into remote access.
What is confirmed and what is not?
- Disclosure to reported hot-fix
- 1 day
- In-the-wild exploits reported (as of Nov 2026)
- 0
- Top Muse bug bounty award (Meta, overall)
- $300k
| Question | Status as of November 2026 |
|---|---|
| Was it fixed? | Wardle reported a hot-fix on 22 September |
| Did Meta comment publicly? | We found no quoted Meta statement |
| Which versions were affected? | Not stated in coverage we found |
| Was it exploited in the wild? | No exploitation reported |
| Was a bounty paid? | Unknown; Meta's Muse bounty offers up to $300,000 overall and $130,000 for single-user prompt injection |
| Does it affect phones? | Disclosure was about the Mac app; other devices were at risk via a compromised Mac |
We will update this page if Meta publishes details. Other Muse incidents are logged on our sister site's 2026 agent incidents page.
Why does this matter beyond one bug?
Three lessons apply to every personal agent, not just Muse.
An agent concentrates access
Muse can act across your email, calendar, shopping and messages. Malware that hijacks the agent inherits that reach without having to break into each account separately. That is why the lethal trifecta and agent takeover protections matter.
Trusted input channels are attack surfaces
Voice dictation feels like you talking to your agent. If something else can speak on that channel, the agent cannot tell the difference. This is a cousin of indirect prompt injection, with the difference that the injected text arrives through a channel Muse treated as the user.
Model safeguards do not fix app bugs
Meta describes extensive defences for Muse, including surrogate tokens so the agent never sees passwords, and says plainly that "prompt injection remains an open problem in the industry." None of those model-level controls help if the app itself can be redirected. Agent safety depends on ordinary software security too.
It is also a reminder that bug bounties and quick fixes work best when researchers report flaws. Meta runs a Muse bounty with awards of up to $300,000, and the turnaround here was fast. Whether this report was paid has not been disclosed.
What should Muse users do now?
- Update the Mac appInstall the latest Muse version, and turn on automatic updates if offered.
- Keep approvals onRequire confirmation for purchases, messages and sharing so an injected instruction still hits a prompt.
- Review recent activityCheck Muse's activity history for tasks you did not start.
- Check connected devicesSign out Muse on devices you no longer use.
- Avoid paste-into-Terminal instructionsNever run commands a website or message tells you to paste.
If you see activity you did not authorise, revoke Muse's connections, change your Meta account password and sign out other sessions. Our guides to Meta Muse safety settings and revoking agent access walk through the steps. If money moved, contact your bank or card issuer and report fraud at reportfraud.ftc.gov.
That guide is at fake AI agent apps, extensions and invites.
Bottom line
The Muse Mac zero-day was serious in principle and short-lived in practice: disclosed on 21 September 2026, reported fixed the next morning, with no reported victims. Its lasting value is as a reminder that an agent's power is only as safe as the device and app it runs on. Keep the app updated, keep approvals on for actions that matter, and treat malware on a Mac running Muse as a threat to everything Muse can touch.



