Meta's Muse AI Agent Freaked Out Its Own Staff Over Leaked Private Messages

technology

Meta's Muse AI agent alarmed its own employees after it repeatedly read private messages and surfaced financial details that appeared to belong to strangers.

Muse, which has been downloaded more than 10 million times since launching in September, began drawing privacy complaints from users, and internal chats revealed Meta staff testing the tool raised the same red flags.

Staff reported the agent reaching into encrypted messages users never intended it to review, mixing up conversations between accounts, and in some cases exposing sensitive data. One employee chatting about fishing received an unprompted response claiming a Capital One balance of $2,722.58 and a Charles Schwab balance of $5,899.57. Another said Muse accidentally rolled their dog into a premium insurance plan charging $888 a month instead of $200, money they could not get back.

Meta dismissed the incidents as isolated and blamed a bug that caused the model to hallucinate messages, insisting the problem was not a security or privacy breach and that it had already been fixed.

Independent testing painted a different picture. On Android, the agent could read Signal messages by leaning on setup permissions instead of Meta's Connector tool, while on macOS it repurposed accessibility features, including screenshot tools, to retrieve messages. Attempts on iPhone were blocked. The checks suggested the issue was ongoing and could be affecting other users.

Signal President Meredith Whittaker accused Meta of launching Muse without respect for basic privacy and urged operating systems to let developers fence off their apps from AI agents. Both Apple and Google are now weighing changes that would give people more control over what they authorize with AI agents, with Apple promising tighter, more explicit consent for Full Disk Access.