Skip to content
The Cybersecurity Focus
The Cybersecurity Focus

Cybersecurity news, threat intelligence, and vulnerability research.

The Cybersecurity Focus

Cybersecurity news, threat intelligence, and vulnerability research.

  • Home
  • Home
The Cybersecurity Focus

Cybersecurity news, threat intelligence, and vulnerability research.

The Cybersecurity Focus

Cybersecurity news, threat intelligence, and vulnerability research.

  • Home
  • Home
exploitMalwareVulnerability

Meta Muse AI Assistant Creates macOS Backdoor Vulnerability

By ogwatermelon
September 23, 2026 4 Min Read
0
September 22, 2026

Security researcher Patrick Wardle has disclosed a serious vulnerability in Meta’s new Muse AI assistant for macOS. The flaw lets malware already on a Mac quietly hijack Muse and abuse the broad permissions its owner granted the app. Wardle published a proof-of-concept on September 21 that demonstrates how an attacker can redirect dictation traffic and steal authentication tokens.

What Happened: Meta Muse macOS Flaw Turns AI Assistant Into a Backdoor

Meta launched Muse in the United States earlier this month. The personal AI agent can access files, email, messages, calendar, shopping, and smart-home apps across Apple devices. Wardle found that Muse exposes an undocumented setting called endo_voyager_dictation_endpoint in its Mac app preferences.

Any program running as the logged-in user can change this endpoint without needing extra permissions. Once modified, when the user taps the microphone and dictates a prompt, the audio and text go to an attacker-controlled server instead of Meta. Wardle warned that the flaw is trivial to exploit and urged people not to install Muse.

This is a local attack. An attacker must already be able to run code as the logged-in user. However, it cannot break into a Mac on its own. Wardle noted that a remote attacker could still pull off the hijack through a ClickFix trick that fools the user into running a single command. The concern is that Muse may have significantly broader access than ordinary local malware, making it a useful target for privilege and access amplification.

Technical Details of the Muse Vulnerability

The vulnerability sits in the Mac version of Muse and stems from how the app handles dictation. Instead of using Apple’s on-device dictation, Meta built its own off-device dictation pipeline that sends audio to a remote endpoint. The endpoint URL is stored in an undocumented preference setting that lacks adequate access controls.

When the endpoint is redirected, Wardle demonstrated three things an attacker can do:

  • Capture dictated audio and prompts sent by the user
  • Inject extra instructions that Muse trusts and acts on
  • Steal the Muse authentication token to control the assistant directly

Because a Muse account can be signed in on multiple devices, an attacker with the stolen token can give orders to Muse on any of them, not just the compromised Mac. Wardle used the token to direct the Muse app on his iPhone to report its exact location, run a Bluetooth scan of nearby devices, and list the smart-home commands it could send.

Security software may not detect the attack because the commands come from Muse, a normal signed app, rather than from something that looks like malware. Wardle also noted that the attack does not defeat macOS protections that stop one app from reading another app’s saved passwords and tokens. It works by getting Muse to act with access it already has.

Business and Operational Impact

The Muse backdoor vulnerability carries significant risk for both individual users and organizations that adopt Meta’s AI assistant. The primary impacts include:

  • Data exfiltration: Attackers can capture voice prompts that may contain sensitive business information, personal credentials, or confidential queries
  • Lateral movement: Stolen tokens enable remote control of Muse across all linked devices, expanding the attack surface beyond the initial Mac
  • Supply chain concern: Muse is a new product from a major platform. Early adoption in enterprise environments could expose organizations before a confirmed patch is available
  • Trust erosion: The flaw undermines confidence in AI assistants that require broad permissions to function

Meta has over 3.9 billion monthly active users across its platforms. Muse is positioned as a personal AI agent with deep integration into daily workflows. A successful exploit could affect millions of users if the assistant sees rapid adoption.

Mitigation and Recommendations

Until Meta confirms a comprehensive fix, Mac users and organizations should take immediate steps to limit exposure.

Immediate Actions for Defenders

  1. Remove or quit Muse on macOS devices until a security advisory is published
  2. Review and revoke unnecessary permissions granted to Muse, including file access, calendar, and smart-home controls
  3. Treat Muse accounts as potentially exposed if the Mac may already be compromised, and change passwords for connected accounts
  4. Avoid voice input in Muse, which closes the specific attack path Wardle demonstrated
  5. Train users not to run commands pasted into Terminal from websites or messages, the typical ClickFix attack vector

Organizational Guidance

Security teams should add Muse to their software inventory and assess whether any endpoints in the environment have the app installed. Consider blocking Muse installation via mobile device management policies until the vulnerability is resolved. Monitor for unusual network traffic to non-Meta domains from Muse processes.

Bottom line: Meta Muse’s undocumented dictation endpoint creates a serious local attack surface. Wardle’s proof-of-concept shows that the AI assistant’s broad access can become an attacker’s broad access. Users and organizations should avoid Muse on macOS until Meta releases a confirmed security fix.

Incident Summary

CVE ID / Incident: Meta Muse macOS Backdoor Vulnerability (no CVE assigned)
Affected Systems: Meta Muse macOS application
Disclosure Date: September 21, 2026
Patch Status: Meta reportedly pushed a partial fix; no security advisory published
Researcher: Patrick Wardle, Objective-See
Attack Vector: Local privilege escalation via undocumented preference setting

References

  1. Wardle, P. (2026, September 21). “not-a-mused: A proof-of-concept for a local Muse vulnerability.” GitHub. https://github.com/pwardle/not-a-mused. Accessed September 22, 2026.
  2. The Hacker News. (2026, September 22). “One Hidden Meta Muse Setting Could Let Attackers Turn the AI Assistant Into a Backdoor.” https://thehackernews.com/2026/09/one-hidden-meta-muse-setting-could-let.html. Accessed September 22, 2026.
  3. PBS NewsHour. (2026, September). “Meta launches personal AI agent Muse to help with everyday tasks.” https://www.pbs.org/newshour/nation/meta-launches-personal-ai-agent-muse-to-help-with-everyday-tasks. Accessed September 22, 2026.

Tags:

ExploitMalwareVulnerability
Author

ogwatermelon

Follow Me
Other Articles
Previous

WordPress CVE Critical Path Traversal Enables RCE

AI Botnet Breach CVE Exploit Hack Incident Infostealer Linux Mac Malware Network Phishing Ransomware RCE supply chain Vishing Vulnerability Windows World Zero Day

Meta

  • Log in
  • Entries feed
  • Comments feed
  • WordPress.org

Archives

  • September 2026
  • August 2026
  • July 2026
  • June 2026
  • May 2026

Categories

  • AI
  • BotNet
  • Breach
  • CVE
  • exploit
  • Hack
  • Incident
  • Infostealer
  • Linux
  • Mac
  • Malware
  • Phishing
  • Ransomware
  • RCE
  • supply-chain
  • Uncategorized
  • Vishing
  • Vulnerability
  • Windows
  • World
  • Zero Day
Copyright 2026 — The Cybersecurity Focus. All rights reserved.