Meta Platforms promised that users of its Muse AI assistant could choose which applications the agent connects to. Yet tests reported found that Muse could retrieve Signal messages without using Meta’s app-connection feature, exposing a gap between selecting connected services and granting wider access to a device.
In its September 8 launch announcement, Meta said users could determine how much access Muse receives. The agent also needs operating-system permissions to view screens, handle notifications, and carry out tasks.
“People choose which apps Muse connects to and exactly how much access it gets,” the tech giant said.
Reuters, working with Signal and the AI Now Institute, tested whether those permissions offered a separate route to messages without using Muse’s dedicated Connector feature.
On Android, the agent retrieved Signal text from push notifications. On macOS, it used accessibility and screenshot capabilities, while equivalent requests were blocked in Reuters’ iPhone tests.
The researchers specifically prompted Muse to read the messages. Their findings do not establish that it bypassed Signal’s end-to-end encryption or that it accessed data without any device-level permission.
The platform distinction has already prompted Apple to introduce more granular controls for app and device access in macOS 27.3 beta 2, including options to limit access to specific applications. The changes are not yet a confirmed general release, and Apple says previously granted permissions remain effective.
Google parent Alphabet told Reuters that Android notification access requires deliberate user authorization.
Meta spokesperson Daniel Roberts called the incidents described by Reuters “isolated incidents” and said the company was improving protections. He maintained that Muse accesses other applications through permissions users explicitly grant.
Reuters reported that Muse had surpassed 10 million downloads since its September launch. Meta’s October earnings release projected $135 billion to $145 billion in 2026 capital expenditures for its broader business and AI infrastructure.
The outlet also examined more than two dozen internal employee reports describing unexpected agent behaviour between September and November. Employees said Muse surfaced banking figures they had not provided, initiated unrelated conversations, and, in one reported case, booked a non-refundable Airbnb stay without authorization.
Privacy concerns had surfaced before this investigation. In September, Inc. columnist Jason Aten reported that Muse referenced private Apple Messages conversations he said he had not authorized it to read. Meta disputed his interpretation.
Signal President Meredith Whittaker called the deployment “technologically reckless” and urged enforceable mechanisms to keep AI agents away from specified applications.