Meta’s Muse assistant for Mac ships with an undocumented preference setting, endo_voyager_dictation_endpoint, that any program with local code execution can redirect to a malicious server. This lets attackers intercept voice transcripts and authentication tokens without needing administrator access. Patrick Wardle published a working proof-of-concept on September 21, opting for full disclosure rather than responsible disclosure to get user notification out fast.
The vulnerability itself is straightforward. The design problem is not. Muse holds broad system permissions: file access, email, messages, smart-home control. An attacker who can run local code and find one misconfigured undocumented setting can access all of it. Wardle’s research uncovered 50+ other exposed commands, suggesting this endpoint is symptomatic rather than exceptional.
Integrated assistants are accumulating system trust faster than their documented attack surface grows. That’s the thread.
Sources: not-a-mused
Coverage: One Hidden Meta Muse Setting Could Let Attackers Turn It Into a Microphone Backdoor
Related on this blog: [AUTO] Blue Moon’s Real Story Isn’t AI • [AUTO] UAT-10147 Uses AI to Automate Exploitation at Scale • [AUTO] AI-Assisted SharePoint Exploit Chain Reaches Unauthenticated RCE