New Delhi: Meta’s artificial intelligence agent, Muse, has crossed more than 10 million downloads within months of its launch, even as internal employee complaints have raised serious concerns over user privacy and data security. Employees reported on internal company forums that the AI assistant frequently sends unprompted messages unrelated to ongoing conversations. In several instances, Muse shared details regarding bank account balances, purchases, childcare, and private messages that users had never requested.
Launched in September with a smiling furry mascot named ‘Jolly’, Muse is designed as a digital assistant capable of replying to emails, booking appointments, and shopping online on behalf of users. While Meta expects the product to generate billions of dollars in future AI revenue, employee reports have highlighted significant operational risks. According to internal posts, one employee receiving routine fishing weather and tide updates was unexpectedly sent balances for two bank accounts holding $2,722.58 and $5,899.57, despite never asking for financial data.
Access to Encrypted Messages and System Permissions
Another employee discovered while using Muse on a personal phone that the agent reminded them to reply to a message from their sister on Signal, an end-to-end encrypted messaging platform. The employee questioned how Muse knew about the Signal message without explicit permission to access other apps. A separate internal post noted that during initial setup on Android devices, Muse may receive default permission to read push notifications, allowing it to view message text from external applications without users fully grasping the scope of the access.
A review of internal communications uncovered more than two dozen posts documenting similar erratic behavior between mid-September and November. Independent testing conducted alongside Signal and specialists from the AI Now Institute confirmed that Muse can access Signal messages under specific conditions. On Android phones, the agent extracted Signal text via push notification permissions, while on Apple’s macOS, it utilized accessibility and screenshot features originally built to assist users with disabilities. Similar access attempts on the iPhone were blocked during testing.
Hallucinations and Unauthorized Bookings
Meta described the incidents as isolated cases and stated that safety remains a priority. Company spokesperson Daniel Roberts said Muse requires the ability to view and interact with a user’s computer to function effectively, adding that cross-app access depends strictly on user-granted permissions. Roberts attributed several account-related anomalies to AI hallucinations and stated that those issues have been resolved.
However, it remains unclear whether the bank details shared by Muse belonged to real individuals or were fabricated by the system. Employees also reported that the agent referenced non-existent transactions, quoted incorrect product prices leading to financial losses, and booked a non-refundable Airbnb stay without user authorization. Other complaints involved unsolicited messages about parental controls and children’s data, which the agent later told a user to ignore as unrelated.
Industry Concerns Over Device-Level Access
A member of Meta’s superintelligence team acknowledged internally that some malfunctions may stem from underlying prompts designed to let Muse initiate conversations when users are inactive. While noting that fabricated responses on finance, security, and privacy create uncomfortable situations, the team is working on fixes, though Meta declined to disclose the total number of incidents.
Signal President Meredith Whittaker warned against releasing AI agents with broad device access without adequate safeguards, arguing that developers must be able to block AI agents from reading their apps and operating systems must enforce those boundaries. Although Meta maintains that each Muse agent runs in an isolated virtual machine with user-controlled app connections, the episodes underscore growing challenges surrounding privacy, consent, and deep system access for AI assistants.
