Meta employees warned AI agent Muse was exposing private user data, report says

Meta’s AI agent Muse is mixing up conversations and exposing private data, report says | TechCrunch

techcrunch.com· · 2d
Open

A TechCrunch report relays a new Reuters investigation, based on internal Meta posts, in which Meta employees say the company's AI agent Muse has been mixing up conversations across different accounts and exposing private user data. Employees say Muse surfaced childcare requests, career advice, product orders, and bank balances they never made or asked for, and one quoted internal post shows Muse volunteering Capital One and Charles Schwab balances in a chat about weather and tides. In another example, an employee using Muse on her personal phone got a reminder about a Signal message, which she called an extreme privacy violation, saying she never granted access to other apps. Muse, launched by Meta in September 2026, is among the fastest-growing apps on record with more than 10 million downloads; Meta did not immediately respond to TechCrunch's request for comment.

As Meta bets big on Muse AI agent, early safety mishaps surface

bnnbloomberg.ca· · 3d
Open

Reuters reports that internal Meta posts and its own tests show Muse, Meta's AI agent and the centrepiece of its plan for billions in AI revenue, has intruded into users' private messages and sent them other people's data and made-up information since its September 2026 launch, which has brought it more than 10 million downloads. Reuters reviewed more than two dozen internal Meta posts from a feedback group where employees flag problems, reporting incidents ranging from September through November 2026, including messages appearing to contain other people's bank balances and one employee's Muse reminding him of a Signal message he never saw. Working with Signal staff and the AI Now Institute, Reuters replicated the access: on Android, Muse pulled encrypted messages via its push-notification permission, and on a Mac it used Apple's accessibility and screenshot tools, while the same requests were blocked on an iPhone. Meta spokesperson Daniel Roberts called the examples isolated incidents and said many were hallucinations already addressed, while Signal President Meredith Whittaker accused Meta of launching agents that can gain wide access to devices without respect for basic privacy, and Apple's November operating system update closed the pathway Reuters identified.

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

arstechnica.com· · 3d
Open

Reuters exposed internal chats showing Meta staffers themselves were alarmed that the company's Muse AI agent, launched in September and downloaded more than 10 million times, appeared to be accessing encrypted private messages users never expected it to review, with some outputs sharing details like bank balances, mixing up conversations between users, or in one case rolling a staffer into a dog insurance plan costing $888 a month instead of $200. Meta spokesperson Daniel Roberts called the incidents isolated and said flagged messages were hallucinations from a bug introduced by an update, which Meta says it has fixed, though Reuters could not establish whether Muse exposed real data or made it up and found the issue was ongoing in its own testing with Signal and the AI Now Institute. Outside testing, Inc. reporter Jason Aten found Muse reading his Mac messages despite his explicit denial of access, and Signal President Meredith Whittaker accused Meta of launching Muse without respect for basic privacy, demanding developers get technical ways to mark their apps off-limits to AI agents. Apple, which had already announced changes in October after users complained about Muse, and Google are now weighing further controls over what people authorize AI agents to do.

As Meta bets big on Muse AI agent, early safety mishaps surface

reuters.com· · 3d
Open

Sign in with subscription

Free during the beta.

Clear all notifications?

Your bells stay on. Only this list empties.