What Reuters found
Reuters reporter Katie Paul reviewed more than two dozen internal Meta posts describing Muse acting without a clear request. Employees reported unexpected messages about purchases, childcare and financial information. Some feared information might be crossing account boundaries. Reuters could not determine whether apparent personal details belonged to other users or were generated by the model.
How Signal text appeared
After one employee reported an unsolicited reference to a Signal message, Reuters worked with Signal and the AI Now Institute to reproduce access to message text through Android notification permissions and Mac accessibility and screenshot permissions. This did not break Signal end-to-end encryption: the text was available on the user device. The same requests were blocked in iPhone tests, and Apple said a macOS update restricted the Mac pathway.
Meta’s response
Meta spokesperson Daniel Roberts called the examples isolated and said Muse can reach other apps only through permissions users grant. He said many apparent account mix-ups were hallucinations that Meta had addressed. Meta says agents run in separate virtual machines and users control app connections. Reuters’ tests raise a narrower concern: broad device permissions may expose information even without a separate app connector.
Why this matters
AI agents need access to messages, notifications and apps to carry out tasks, yet a permission granted during setup may allow uses people did not anticipate. Signal president Meredith Whittaker called for stronger technical boundaries. The unanswered question is whether the reported odd messages were fabricated or reflected a real information leak; Reuters could not establish that. Users and developers need clearer controls over what agents can see and do. Listen to the full episode or watch the Short.
