The honest view from inside technical leadership
Analysis

Engineering leadership

Independent reporting on the decisions, costs and contracts that shape how engineering organizations build with AI.

Agents · News

Meta’s Muse AI Agent Showed Employees Bank Balances and Orders That Belonged to Other People

Internal posts reviewed by Reuters describe Meta’s Muse AI agent placing other people’s messages, purchases and account balances in front of the employees testing it. Meta says employee testing is how it finds and fixes privacy problems before features reach the public.

Line-art illustration of two chat windows with a bank balance card crossing from one into the other, representing the Meta Muse AI agent showing employees other people’s private data

Meta’s Muse AI agent showed some Meta employees private information that belonged to other people, including bank account balances, product orders and childcare requests they had never made, according to internal company posts reported by Reuters on 18 November 2026. Meta says the employees were testing versions of Muse that differ from the app the public uses.

Meta Employees Reported the Muse AI Agent Mixing Up Conversations Between Accounts

Employees testing the Meta Muse AI agent wrote in internal posts that it was blending conversations from different accounts, Reuters reported, in an account that TechCrunch’s Aisha Malik summarized on 19 November 2026. The examples were specific. Employees described Muse surfacing childcare requests they had not made, career advice they had not asked for, orders for products they had not bought and bank balances that did not belong to them. Muse is Meta’s personal AI agent. It launched in September 2026, and TechCrunch reports it has passed 10 million downloads.

A Weather Question Returned Two Bank Balances Nobody Had Asked For

One employee told colleagues that Muse, while answering a chat about weather and tide information, added an extra message nobody had requested. It listed a Capital One balance of $2,722.58 and a Charles Schwab balance of $5,899.57. In a second case, an employee using Muse on a personal phone received a reminder from the agent to reply to a message from a sister on Signal, the end-to-end encrypted messaging app. The employee wrote that they had not given Muse access to other apps and did not remember being asked. “This seems like an extreme privacy violation,” the employee wrote, according to the report.

Cross-Account Data in an AI Agent Comes From the Retrieval Layer Around the Model

Meta has not said what caused the mix-ups, and the reporting does not identify a cause. What is known is where such data has to travel. A language model can only repeat information that the software around it has placed into its context: stored memories, past conversations, and data pulled from connected apps such as banking or messaging. For one person’s balance to appear in another person’s chat, some step that gathered context had to fetch it under the wrong identity or from a store that more than one account could reach. Both are tenancy failures of the kind that databases and caches have suffered for decades. An agent adds two complications. The model may volunteer the data unprompted, as the weather answer did, and the data arrives in fluent prose that looks like a normal reply, so a tester may read past it.

Meta Says Employees Test Versions of Muse the Public Does Not Have

A Meta spokesperson told TechCrunch that employee testing is central to how the company develops products, and that its purpose is to gather feedback so Meta can build safety and privacy protections. The spokesperson added that employees have access to different versions of Muse than the public and often review features and behavior that are not available anywhere else. TechCrunch updated its story at 12:05 PM Pacific on 19 November to include the comment. The statement does not say whether the public version of Muse can mix data between accounts, whether any public user was affected, or whether the problems employees reported have been fixed.

Why This Matters for Engineering Leadership

Any team shipping an agent with memory or connected apps is running a multi-tenant data path, and every piece of context the model sees is a read from that path. The isolation has to be enforced in code before the model receives anything: each retrieval, memory lookup and connector call keyed to the authenticated user, with no shared cache or index that can return another user’s rows. The test for it is cheap. Seed several test accounts with unique marker data, such as fake balances or invented names, run many sessions at once, and fail the build if any account’s markers appear in another account’s replies, logs or memory. Then record, for every response, which sources contributed to its context, so that a report like the weather message can be traced to the exact retrieval that caused it. The Signal reminder points at the same discipline from the user’s side: an agent should be able to show each user which apps and data it can read, because the employee who saw it had no idea that access existed. Meta’s employees found these problems in internal testing, and internal testing is where a cross-account leak should be found.

FAQ

What did Meta’s Muse AI agent show employees?

According to internal posts reported by Reuters, Muse showed some employees childcare requests, career advice, product orders and bank balances that were not theirs. Employees described it as mixing up conversations across different accounts.

Were members of the public affected by the Muse data exposure?

The reported cases come from Meta employees testing Muse. Meta says employees use different versions of Muse from the public and has not said whether the public app has the same problem.

What has Meta said about the Muse privacy reports?

A Meta spokesperson told TechCrunch that employee testing is core to product development and exists to gather feedback for safety and privacy protections. The spokesperson said employees often test features the public does not have.

What caused Muse to mix up conversations between accounts?

Meta has not given a cause, and the reporting does not identify one. In any AI agent, data reaches the model through retrieval and memory steps that must be scoped to the signed-in user.

How many people use Meta’s Muse AI agent?

TechCrunch reports that Muse, launched in September 2026, has passed 10 million downloads.

How can engineering teams test an AI agent for cross-account data leaks?

Seed separate test accounts with unique marker data, run many sessions in parallel, and fail the build if one account’s markers ever appear in another account’s responses or logs.

Source: Meta’s AI agent Muse is mixing up conversations and could be exposing private data, report says by Aisha Malik, TechCrunch, November 19, 2026. Original reporting: Reuters investigation into Meta’s Muse AI agent, November 18, 2026.

Read more news & analysis