Un-Mused: How a Single Debug Setting Bypassed macOS Security in Meta’s AI Client
Security researcher Patrick Wardle revealed an unpatched zero-day vulnerability in Meta's Muse desktop client for macOS. This flaw lets unprivileged software manipulate the assistant's extensive permissions, compromising input confidentiality and account security. Despite a hotfix from Meta, the vulnerability raises significant concerns about platform trust and security boundaries. By Olimpiu Pop
Security researcher Patrick Wardle revealed an unpatched zero-day vulnerability in Meta's new desktop Muse client for macOS. Despite Meta's CEO Mark Zuckerberg's claim that the AI assistant was designed with privacy and security as top priorities, the flaw allows local software to bypass macOS security measures. By manipulating an undocumented debug setting, unprivileged software can hijack the application and circumvent standard macOS security boundaries.
This vulnerability, named endo_voyager_dictation_endpoint, enables attackers to redirect dictation traffic to a server under their control, potentially capturing authentication tokens and audio data. The flaw also grants attackers access to prompt injection attacks, allowing them to force the assistant into performing unauthorized tasks.
Since Meta did not release a security advisory or coordinate with a CVE Numbering Authority, the vulnerability currently lacks an official CVE designation. Meta quickly released a hotfix to remove the internal debugging preference from production client builds, effectively resolving the issue.
Written by urgent.news from InfoQ's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
