Key points
- Reuters and collaborators reproduced Signal message access via broad device permissions.
- Meta says many unusual messages were hallucinations; cross-account leakage remains unproven.
- The tests did not reproduce the behavior on iPhone.
Meta's consumer AI agent Muse is facing renewed scrutiny after a Reuters investigation published on November 18 documented employee complaints about unexpected actions and reproduced a route through which the agent could read messages from Signal, an encrypted messaging app. The findings raise questions about the boundaries between permissions granted to an operating system and access users believe they have authorized for a particular AI service.
What the investigation established
Reuters reviewed more than two dozen internal reports describing unwanted Muse behavior. Employees described unsolicited replies referring to unfamiliar financial balances, shopping arrangements and family matters. One worker reported an Airbnb booking made without permission. These are accounts from internal feedback, not independently established proof that the agent transferred another person's private data.
The stronger technical finding concerns Signal. Working with Signal staff and the AI Now Institute, Reuters reproduced access to Signal message text on Android through notification permissions and on macOS through accessibility and screenshot capabilities. Users had not necessarily linked Signal using Muse's dedicated connector. Reuters said the same tests did not reproduce that behavior on an iPhone.
End-to-end encryption protects messages in transit and against access by the messaging service, but it does not automatically prevent software with sufficient access to an unlocked device from reading content displayed or delivered in notifications. The distinction is central: the investigation does not demonstrate that Signal's encryption was broken.
Meta's response and the unresolved question
Meta spokesperson Daniel Roberts called the examples isolated incidents and said the company conducts testing and updates safeguards. He said Muse requires device permissions to work on users' behalf and that many seemingly mismatched conversations were AI-generated fabrications, commonly called hallucinations.
Reuters could not establish whether the unfamiliar balances and other sensitive details actually belonged to different users or were invented by the model. Consequently, it would be premature to characterize every example as a confirmed cross-account data breach. The reproduced access to message text, however, is a separate and concrete concern about permission scope.
Why device permissions matter
Personal agents promise to book services, send messages and navigate apps. To perform those tasks, they may request broad notification, accessibility or screen-reading privileges. Those privileges can expose information beyond the specific service a user intended the agent to handle. An app-level connection setting is not necessarily the same as an operating-system permission.
The Reuters report also describes how Apple tightened a macOS access pathway, while platform makers considered additional controls. The broader policy question is whether agent providers should require more granular, understandable consent before accessing communications and acting on behalf of users.
What users and regulators should watch
Users can review notification and accessibility permissions, disconnect unnecessary integrations and require explicit confirmation for purchases or bookings. These steps reduce exposure but cannot guarantee error-free agent behavior.
The immediate unresolved issues are the prevalence of the reported failures, whether any genuine cross-account disclosure occurred, and whether revised controls prevent recurrence. Meta's explanations and the independent test findings should be evaluated separately rather than treating either as a complete account of every incident.




