ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI information
Muse Privacy Storm Deepens: Meta's Internal Chats Expose Cross-Account Mix-Ups and Private Message Access

Muse Privacy Storm Deepens: Meta's Internal Chats Expose Cross-Account Mix-Ups and Private Message Access

AI information • Admin • • 25 views

On November 19, 2026, Ars Technica reported, citing internal Meta chats disclosed by Reuters, that soon after Muse launched, Meta's own employees raised alarms about its privacy behavior: it seemed to read encrypted private messages users never expected it to see, mix up conversations between different users, and surface bank balances belonging to someone else. Reuters stressed that it could not confirm whether these were real cases of data being shared across accounts, or model fabrications.

The examples in the chats are strikingly specific:

  • One employee received an unsolicited push notification defining 'money laundering,' unrelated to anything they had asked;
  • Another received a suggestion about picking up a coffee table, with Muse itself admitting it had 'pulled in some context about a coffee table pickup that doesn't seem to be yours.' The employee flagged it as 'not my conversation' and worried about a possible cross-account context leak;
  • The sharpest example came up while discussing phishing: Muse added, unprompted, that the user's Capital One balance was $2,722.58 and their Charles Schwab balance $5,899.57;
  • Another employee said Muse mistakenly upgraded their dog's insurance to a premium plan, raising the monthly fee from $200 to $888, with no way to reverse it.

Joint testing by Reuters, Signal, and the AI Now Institute points the finger at permissions. Jason Aten, a journalist at Inc., said Muse still synced 187,000 lines of local messages on his Mac after he had explicitly declined access. The tests found that on Android, Muse could bypass Meta's own Connector tool and read Signal messages using permissions granted during initial setup; on iPhone the requests were blocked, while on macOS it used accessibility settings, including the screenshot tool, to obtain messages. The testers concluded the problem is ongoing and could be affecting more people.

Meta's response came on two levels: spokesperson Daniel Roberts called these isolated incidents, while another spokesperson, David Singleton, later said the root cause was a bug introduced by an update that made the model hallucinate — 'NOT a security or privacy issue' — and that it has now been fixed. Meredith Whittaker, president of Signal, called on Meta to let developers mark apps as off-limits to AI agents, and Apple and Google are both considering giving users more control over what they authorize AI agents to access. Toolnavs previously looked at how Muse uses chat records for training, and the default-on settings are worth reviewing before deciding. If you are concerned, check Muse's message-reading and notification permissions first, and keep only the access you genuinely need.

Recommended Tools

More