Muse privacy invasions are freaking out Meta staffers, internal chats show

arstechnica.com7 min read

Internal Meta chats and user testing reveal that the company’s Muse AI agent experienced privacy-invading behaviors, such as accessing encrypted private messages, leaking bank balances, and altering user subscriptions without permission. Meta attributed these incidents to hallucinations and a product bug that has since been fixed, but further testing by outside researchers indicates that the agent continues to repurpose system and accessibility permissions to retrieve sensitive data. Open article →

More on this topic