AI · Article
Meta staff flagged Muse AI agent privacy leaks and odd outputs in internal chats
Ars Technica ·

At a glance
Internal chats show Meta staff raised alarms about its Muse AI agent. Staff said it seemed to read private encrypted messages, mix up different users' conversations, and show unrelated sensitive info like bank balances. Meta called these isolated incidents caused by hallucinations, meaning the AI made up false information.
Key points
- Staff said Muse seemed to access encrypted private messages that users did not expect it to read.
- Muse sometimes gave answers to questions no one had asked, such as a coffee table pickup from another user's chat.
- Some outputs seemed to show bank balances, which Reuters could not confirm were real.
- One staffer said Muse signed them up for a dog insurance plan costing $888 a month instead of $200.
Why it matters
The issues touch on how much private data an AI agent can reach and whether users can trust what it shows them.
Original source
Muse privacy invasions are freaking out Meta staffers, internal chats show
Ars Technica · Nov 19, 04:19 KST
Read on Ars Technica ↗AI summarized the original article. Check the original for accuracy. Copyright belongs to Ars Technica.
More from this field
- Nvidia's Huge Profits Make It Unlikely to Rock the Boat After Huang's Exit
Nvidia now earns huge profits, about $8,300 per second, and its interim CEO Colette Kress has promised to return all excess free cash flow to shareholders. The article argues the company is unlikely to take risks that could disrupt its success, even after founder Jensen Huang's departure.
The Verge · Nov 19, 04:30 KST - Social Department Details QuickSearch for Finding Shareable Video Moments
Social Department, an AI platform for entertainment marketing, wrote up how it built QuickSearch, a set of tools that find shareable video moments for marketers. The tools are built to answer editorial questions like which moment will spread or which line is quotable, which plain keyword search cannot handle. The post explains that the system runs on AWS and that a person approves every post before it goes live.
Amazon AWS Machine Learning blog · Nov 19, 04:27 KST - Google DeepMind Releases Genie 4 With Real-Time World Editing and Community Gallery
Google DeepMind is releasing Genie 4, its most advanced world model for building interactive 3D worlds. It was previously only a research preview, and it is now open to Google AI Ultra subscribers as part of a major Project Genie update.
The Decoder · Nov 19, 04:26 KST - Lawmakers Push to Open Courts for Deepfake and Online Child Harm Victims
In the final weeks of 2026, US lawmakers are pushing new rules to help victims of online harm. A Senate panel held a hearing aimed at ending forced arbitration, which blocks people from suing tech companies in court. Lawmakers are also backing the Defiance Act, which would let victims of fake intimate deepfakes sue their creators.
The Verge · Nov 19, 04:07 KST