News

Policy & safetyNov 20, 20263 min read1 source

Meta’s AI agent Muse reportedly mixed up users’ messages and bank balances

Worth watching

Meta employees said the Muse AI agent showed them other people’s private messages and bank data, a warning for any business giving an AI agent shared access.

Read it in

Drafted with AI· Nov 20, 2026

Policy & safety · Nov 20, 2026

The distilled version

  • Meta employees reported that its Muse AI agent crossed wires between accounts, showing them other people’s private messages, orders and bank balances, per a Reuters report covered by TechCrunch.
  • It shows how easily an AI agent with access to shared or multiple accounts can leak one person’s private data to another, including financial details.
  • Before giving any AI agent access to shared inboxes, accounts or messaging apps, test it with fake data first and check exactly what it can see.

What happened

According to a Reuters report covered by TechCrunch, Meta employees raised concerns internally that Muse, the company’s AI agent, has been crossing wires between different accounts and exposing private data. Employees said the agent showed them things that belonged to other people entirely: requests about someone else’s childcare, career advice meant for another account, orders that person never placed, and bank balances that weren’t theirs.

In one case, an employee said Muse randomly inserted a message into an unrelated chat about weather and tide information, reading out two bank balances, a Capital One figure of $2722.58 and a Charles Schwab figure of $5899.57, that didn’t belong to that person. In another, an employee using Muse on her personal phone got a reminder to reply to a message from her sister on Signal, and said she never recalled granting the agent access to other apps.

A Meta spokesperson told TechCrunch that employee testing is meant to surface issues like this so the company can add safety and privacy protections, and noted that employees use different versions of Muse than the public, so they see features and behavior not available elsewhere. Muse launched in September and has passed 10 million downloads, making it one of the fastest-growing apps on record.

Why it matters

This is a real example of what can go wrong when an AI agent has access to more than one person’s accounts or messages: it can mix them up and expose private details, including financial ones, to the wrong person.

If you’re considering any AI agent that touches shared email, shared accounts or messaging apps in your business, this is a reminder to find out exactly what it can see and whether it keeps each person’s data separate before you roll it out.

What it means for your business

If you run the business: before connecting any AI agent to shared accounts, email or messaging apps, ask the vendor directly what data it can access and whether one person’s information can ever surface to another.

What to do

Worth watching
  • List which AI agents or assistants in your business already have access to shared inboxes, accounts or messaging apps.
  • Check each tool’s settings for exactly what data it can read, and turn off any access you don’t need.
  • Before rolling out a new AI agent, test it on dummy accounts with fake data, not real customer or financial information.
  • Ask any vendor in writing whether their agent keeps each user’s data separate from everyone else’s.

Worth watching: no action needed yet; revisit when it matures.

Hype check

The claim: Meta says employee testing is meant to surface issues so it can add safety and privacy protections.

The reality: The reports describe real bank balances and personal messages being exposed to employees during that same testing, after Muse had already launched publicly and passed 10 million downloads.

Jargon in this story

AgentPIIGuardrailsData retention

Sources

  1. 1Meta’s AI agent Muse is mixing up conversations and could be exposing private data, report says (opens in a new tab) · TechCrunch · News report · Nov 19, 2026

Keep going: learn it, then do it.