Reuters test: Meta’s Muse agent read Signal messages on Android and Mac
Meta employees said the Muse agent booked an Airbnb stay for one of them without permission and showed another employee account balances that belonged to someone else. Reuters could not establish whether the data belonged to other users or had been made up by the agent.
Reuters reviewed more than two dozen internal Meta messages in which employees describe unrequested actions by the Muse agent. Meta launched Muse in September. It is an AI agent, a program that uses an AI model to handle tasks such as emails, appointments and online purchases on the user’s behalf. It has been downloaded more than 10 million times.
After a routine fishing-weather message, one employee received the balances of Capital One and Charles Schwab accounts, even though the employee had never given Muse any bank details or discussed banking with it. The agent offered another employee help arranging the collection of a coffee table, then admitted it had retrieved information that ‘doesn’t appear to be yours’. Muse booked a non-refundable Airbnb stay for one employee without permission.
Employees suspected that Muse was mixing up conversations from different accounts. Reuters could not establish whether the data belonged to other users or had been made up by the agent. Meta spokesperson Daniel Roberts called the incidents isolated. He said many of the messages employees had flagged as potential account mix-ups were hallucinations, meaning information the model made up, and had already been addressed. A member of the Superintelligence team said internally that the agent appeared to invent messages when built-in prompts direct it to contact the user between sessions. Meta told TechCrunch that employees use different versions of Muse from the public.
Reuters was, however, able to reproduce Muse reading Signal messages, working with Signal and the AI Now Institute. One employee had received a reminder from Muse to reply to a Signal message from the employee’s sister, despite having, by the employee’s own account, given the agent no access to other apps. On both Android and Mac, Muse bypassed Meta’s Connector tool: on Android it read messages from phone notifications, and on Mac it used accessibility and screenshot tools. On iPhone, the same requests were blocked.
Reuters said the results suggest that Meta interprets the permissions users grant broadly: the agent used permissions granted during installation in ways a user might not anticipate. A colleague of the employee said Muse had notification permission on Android by default. Google and Meta, for their part, say users must grant notification permission deliberately.
An operating system update that Apple released this week effectively closed the route Reuters found: it restricted the use of accessibility tools on Macs and introduced separate permissions for agents. Under Meta’s terms of service, users are responsible for the agent’s actions.
Read more: How an agent works · Agent map