As Meta bets big on Muse AI agent, early safety mishaps surface
META•Meta’s Muse AI agent has sent unsolicited chats, reportedly accessed Signal messages through device permissions and provided fabricated purchase and financial information, according to internal posts and tests. The app has more than 10 million downloads since its September launch; Meta described the incidents as isolated and said it continuously updates protections.
1. Reports of unexpected activity
Meta employees reported that Muse sent messages about bank balances, childcare requests, purchases and other topics they had not raised. Some worried that the messages could indicate conversations or data were being mixed between accounts. Meta said many flagged messages were hallucinations and had been addressed, but Reuters could not establish whether sensitive information was exposed or fabricated.
2. Tests find access pathways
Reuters worked with Signal staff and the AI Now Institute to test access to Signal messages. On Android, Muse used permission to read push notifications; on a Mac, it used accessibility and screenshot tools. Tests on an iPhone blocked the same requests. Apple said an operating-system update restricted the Mac pathway, while Apple and Google said they were considering changes to give people more control over permissions.
3. Meta’s response
Meta said Muse can access other apps only through permissions users grant and that it isolates each agent in a secure virtual machine. The company said it prioritizes safety, conducts internal testing and updates products to improve protections. Muse has had more than 10 million downloads since its September launch, and Meta has disclosed that around a quarter of users who set up accounts return daily.




