Reuters reviewed more than two dozen internal Meta posts about its Muse assistant, an AI agent that acts for users. Employees reported receiving bank balances and messages that seemed to belong to other people. Meta called the cases isolated and said many were invented by the AI, not real leaks.