Home / News

Meta's Muse agent reportedly mixed up users' private data: a lesson for business bots

An AI agent that mixes up conversations is a data leak, not a glitch. Anyone running a customer-facing support or sales bot should check one thing now: whether client A can ever see anything that belongs to client B.

TechCrunch, citing a Reuters report, says Meta employees posted internally about the Muse agent showing them childcare requests, product orders and bank balances that weren't theirs. One got a reply about tides with someone's Capital One and Charles Schwab balances added. Another got a reminder to answer a Signal message, though she says she never gave Muse access to other apps. Meta says employees test versions the public doesn't get. We don't know the cause, or whether the public app is affected.

What this means for your own agent

A small business agent handles leads in WhatsApp, bookings, invoices. Cross-client leaks usually come from shared memory, a shared cache or a knowledge base searched without a per-client filter. In a booking assistant like the one for a fitness studio network, isolation has to live in code: tie each session to a customer ID and filter memory and search results by that ID. Give tools only the access they need. Don't rely on a prompt that says "never reveal other users' data".

Why it matters: Before launch, run a simple test with two fake customers and try to pull one's details from the other's chat. If you can't explain how your agent separates customers, it probably doesn't.

Source: TechCrunch