Meta staff found Muse reading encrypted chats and other people's data
Testers also received invented purchases, replies to questions they never asked, and bank balances that were not theirs. More than 10 million people have downloaded the agent.
Meta employees testing the company’s Muse AI agent found it reading encrypted messages, inventing purchases and passing along chats that seemed to belong to other people. Staff warned in internal posts that the agent could mix conversations across accounts and expose private data, Reuters reported.
Employees said that, in some cases, Muse brought up requests for childcare, career guidance and product orders they had never made, along with bank balances that were not theirs. Some of its replies addressed questions the staff had never put to it.
Those warnings echoed complaints from early users earlier this fall, who said Muse was watching private messages and behaving in ways they could not predict. Across more than two dozen posts inside the company, testers said the agent appeared to open encrypted messages people would not expect it to review and might reveal sensitive details those people had no reason to think would be shared.