theaisafemode.com
AI Safe Mode

Staff Testing a New AI Agent Saw It Read Private Chats and Mix Up Accounts

A hand holding a smartphone among blurred coloured lights; illustrative
Illustrative photograph only, not an event image. Photo by Rodion Kutsaev via Wikimedia Commons; CC0 1.0. See LICENSES.md.

Internal posts show employees testing a popular AI agent saw it read encrypted chats they never connected and show what looked like other people's bank balances.

Employees testing a popular new AI agent from a major tech company warned that it read private messages they never meant it to see and mixed up conversations between accounts. Some replies even appeared to show other people’s bank balances, according to internal posts reviewed by a news agency.[1] [2]

What happened

One report said a staff member was reminded by the agent to answer her sister on an encrypted messaging app, even though she said she had never given it access to other apps.[2] Another was offered help messaging a seller about furniture and was told the agent had pulled in context that “doesn’t seem to be yours.”[1] [2] A chat about fishing ended with an unprompted line listing two bank balances.[1] [2]

Staff also found the agent had permission by default to read phone notifications, which can include message previews.[1] [2] Testers working with the news agency were able to repeat the message access on some phones and computers.[1] [2]

The company called these “isolated incidents” and said many suspected mix-ups were AI “hallucinations” it had already fixed. The news agency could not establish whether the agent was showing real data from other users or making it up.[1] [2] The agent has been downloaded more than 10 million times since it launched in September.[1] [2]

Why it matters

AI agents are designed to act across your apps, so a single permission can open far more than you expect. Even if the bank balances were invented, an assistant that confidently shows money or private messages out of context can cause panic, bad decisions or real leaks.

The short version

Staff testing a major AI agent saw it read private chats and mix up accounts. Give agents the smallest access possible.

How to stay safe

  • Review your agent’s permissions for messages, notifications, screen recording and accessibility.
  • Turn off notification access unless a task truly needs it.
  • Never act on money information an AI shows you unprompted. Check your bank directly.
  • Require approval before payments, bookings or messages sent on your behalf.
  • Remove agents you no longer use and revoke their connected accounts.

Follow AI Safe Mode

See the post on YouTube Watch and follow AI Safe Mode on YouTube @AISafeMode and Instagram @aisafemode.

References

  1. Ars Technica, Nov 18, 2026. Report on employees' internal warnings about the AI agent. Read the report ↗
  2. TNW, Nov 20, 2026. Report on the internal posts and the replication tests. Read the report ↗
Facts come only from the published reports above and are attributed to them. AI Safe Mode doesn't name companies, products or people involved in incidents. Spotted an error? Email [email protected] and it will be reviewed and corrected. See the corrections policy.