Meta’s recently launched AI assistant Muse, designed to automate tasks like booking appointments, filling forms, and making purchases, contains a critical zero-day vulnerability that allows complete system compromise. The flaw enables any locally installed application or terminal command to gain unauthorized access to authentication tokens that control user accounts, despite Apple’s safeguards designed to prevent such access.
The vulnerability was discovered by macOS security researcher Patrick Wardle, who identified that Muse’s design allows processes to change settings determining where voice transcription occurs. By redirecting the transcription endpoint to an attacker-controlled server, malicious actors can intercept authentication tokens and gain full control over the compromised Muse account. Wardle demonstrated proof-of-concept attacks capable of writing files to disk and capturing images without user notification.
The flaw stems from several design decisions made by Meta developers. The choice to process dictation in Meta’s cloud servers rather than locally on the device, combined with allowing any application to modify transcription endpoints, created the security gap. Security experts questioned whether adequate consideration was given to security during development, particularly given Muse’s extensive access to user data, accounts, and device resources including cameras, microphones, and file systems.
Meta released a hotfix addressing the vulnerability more than 12 hours after the disclosure. The discovery occurred just before Amazon announced it was blocking Muse from its shopping platform, citing concerns about unauthorized third-party agents making purchases on behalf of customers. Amazon stated that such applications should operate transparently and respect business decisions about participation, comparing Muse to regulated services like food delivery and travel booking apps.
Article Attribution | Read More at Article Source
Article summary produced by Claude AI