Meta's Muse AI Agent Reportedly Mixed Up Users' Private Data

Meta employees raised concerns that Muse, the company's AI agent, was mixing up conversations across different user accounts and exposing private data, according to a report published on 19 November 2026 by TechCrunch. Reuters reported the privacy concerns separately, based on a review of internal company posts, in an investigation of Meta's early safety problems with the product.
According to employees quoted in the reporting, Muse surfaced childcare requests that users had never made, career advice nobody had asked for, product orders that were never placed, and bank account balances that did not belong to the person viewing them. The failure is not a wrong answer in isolation. It is the agent drawing on one user's data and presenting it in another user's session, the chat window a person is currently using.
One example reported by Reuters involved a chat about weather and tide information. Muse added a message listing a Capital One balance of $2722.58 and a Charles Schwab balance of $5899.57 to that conversation. Neither balance belonged to the participant in that chat.
A second example involved an employee's personal phone. Muse reminded her to respond to a message from her sister on Signal. She described the reminder as an extreme privacy violation and said she had not authorized it.
Muse launched in September and has more than 10 million downloads, according to TechCrunch. The scale matters for the nature of the defect. A cross-session leak that touches a small fraction of users is a bug. At ten million installs, even a low incidence rate means a large number of people whose financial and personal details have appeared on someone else's screen.
In my view, the mechanism the employees describe is what separates this from ordinary model error. A hallucinated fact can be checked and discarded. Personal data that appears in the wrong account is a tenant isolation problem, the kind that zero-trust designs and per-user context boundaries exist to prevent. Whether the cause lies in retrieval, memory, session routing or the agent's permission scope is not something the reporting resolves, and it would be premature to assign it.
The broader context here is that agentic products, which take actions and carry state across tools and accounts, widen the surface on which isolation failures can occur. A chatbot that answers only from its own conversation has a narrow blast radius. An agent with access to banking, messaging and purchasing signals has a much wider one. Muse's early behaviour exposed that breadth before the product's safeguards caught up. Whether the company's fixes close the gap is a question the coming months will answer.
For teams building or evaluating similar systems, the practical lesson is concrete. Per-user isolation has to be verified at the level of what the agent can retrieve and what it can emit, not only at the level of what the interface shows. Employee testing surfaced these failures during the product's early period. That is where the more hopeful reading lies: internal scrutiny caught the problem while the product was still being shaped, and a defect that is visible can be fixed. The harder question is whether the testing process catches the next one before it reaches users.


