A security vulnerability discovered in Meta’s newly launched Muse AI agent has highlighted a growing concern around agentic AI: when an autonomous assistant receives broad access to emails, files, payments and connected services, a security flaw can potentially expose far more than a conventional application.
According to reports, an external security researcher discovered a vulnerability that could have enabled unauthorized access to the dedicated virtual machines assigned to Muse users. Meta reportedly classified the vulnerability as SEV-2, its third-highest severity category.
Muse is designed to perform real-world tasks—including shopping, travel bookings, emails and payments—on behalf of users. Dedicated virtual environments give the agent the persistent computing capability required to execute these tasks, but they also create a potentially valuable target for attackers.
The concern is particularly significant because Muse reportedly attracted millions of downloads shortly after launch. Although there is no confirmed evidence that the vulnerability was exploited maliciously, a security weakness at an early stage can undermine user confidence in autonomous AI.
A separate vulnerability involving Muse on macOS reportedly demonstrated another potential attack path. Researcher Patrick Wardle found that malware already running locally could manipulate an undocumented configuration setting and redirect voice-dictation traffic, potentially exposing authentication information and accessing resources available through the application’s permissions. Meta subsequently issued a fix.
