Testers inside Meta saw their AI agent Muse bring up Signal messages, on Android through the phone's notifications
Reuters reviewed internal posts by employees, while Meta speaks of isolated incidents and blames many errors on answers the model made up.
Reuters has seen more than twenty internal posts by Meta employees who tested the AI agent Muse after its launch in September. Among other things, they describe Muse reminding a user to reply to a Signal message from her sister, even though Signal had not been connected as a service. One employee called it an extreme privacy violation, Gadget Review reports, based on the investigation.
Together with Signal staff and the AI Now Institute, Reuters examined how this was possible. On Android, Muse read the messages through the notification access users grant during setup, on a Mac through accessibility features and screenshots. On an iPhone the same request was blocked, Startup Fortune writes. According to Martin Cid Magazine, Apple has since released a macOS update that closes that route and introduces separate permissions for agents, while Google has not announced a similar change.
Testers also reported other problems. Muse showed unrequested balances of accounts at Capital One and Charles Schwab during a conversation about fishing, cited data that appeared to come from another account and enrolled an employee in a dog insurance plan at 888 dollars a month instead of 200 dollars. Reuters could not verify whether the balances were real.
Meta spokesperson Daniel Roberts says Muse reaches other apps only through permissions users grant themselves, and that many apparent account mix-ups were model hallucinations that have since been fixed. Signal president Meredith Whittaker believes Meta launched the agent without respect for basic privacy principles.