AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Un-Mused: How a Single Debug Setting Bypassed macOS Security in Meta’s AI Client

Collected Sep 30, 2026

Security researcher Patrick Wardle, founder of the Objective-See Foundation, disclosed an unpatched zero-day vulnerability affecting Meta's newly released Muse desktop client for macOS, according to InfoQ. The flaw stems from an undocumented configuration preference key named endo_voyager_dictation_endpoint. On macOS, local processes and arbitrary scripts running in an unprivileged user context can overwrite this value without elevated administrative rights or operating system authorization prompts, the report states.

Under normal operation, the parameter designates the cloud endpoint receiving voice dictation audio and returning transcriptions. By modifying it, an attacker can silently reroute the assistant's outbound dictation traffic to a server they control, the report says. When a user activates dictation, the client sends raw microphone audio and a valid authentication token for the victim's Muse account to the configured endpoint. Wardle demonstrated a proxy server capturing tokens and audio while forwarding legitimate traffic to Meta's servers to avoid detection, and noted that prompt injection could append hidden instructions to voice requests.

Wardle published a proof-of-concept exploit titled not-a-mused. The vulnerability lacks an official CVE designation because Meta did not release a formal security advisory or coordinate with a CVE Numbering Authority, the report says.

After disclosure, Meta deployed a hotfix to the macOS Muse application, stripping the internal debugging preference setting from production builds and preventing local modification of the dictation server destination. David Singleton of Meta Superintelligence Labs framed the issue as a local configuration problem requiring prior code execution, and the engineering team resolved it by removing the preference key from production builds, according to the report.

The report also says security professionals disputed that characterization, arguing initial access is trivialized by social engineering lures such as ClickFix. It adds that Amazon moved to block Muse from its shopping platform, citing non-compliance with automated agent access policies. A former Meta artificial intelligence security engineering manager said they would avoid using the software due to risks of deep integration.

Read at InfoQ · AI, ML & Data Engineering

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

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