Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableTechCrunch→ original

Meta's Muse agent mixes up other people's chats and shows employees bank balances

Meta's Muse agent, launched in September, mixes up conversations between different accounts, according to Reuters. Employees wrote in internal posts that it sent them messages about orders and career advice they had not asked for. In the middle of a conversation about the weather, one of them was given bank account balances of $2722.58 and $5899.57. The takeaway for agent builders: data isolation between users needs to be tested as rigorously as access permissions.

AI-processed from TechCrunch; edited by Hamidun News
Meta's Muse agent mixes up other people's chats and shows employees bank balances
Source: TechCrunch. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTechCrunch ↗
Editor-in-Chief
Key metric #1Muse launched in September and has passed 10 million downloads
Key metric #2Muse reminded one employee to reply to her sister on Signal, though she had never granted access
Key metric #3Meta has not yet responded to TechCrunch's request for comment, and the cause of the failure has not been disclosed

Imagine asking an assistant about the weather and tides, and it adds at the end: "By the way, your Capital One account has $2722.58, and your Charles Schwab account has $5899.57." The problem is that these are not your accounts. This is the story a Meta employee described in an internal post, which Reuters cites. The subject is Muse, an AI agent the company launched in September.

What happened

Reuters reviewed internal messages from Meta employees. They complained that Muse mixes conversations from different accounts and shows other people's personal data. According to them, the agent sent messages about requests to find a babysitter that no one had made. It gave career advice nobody had requested and reported orders for goods that they had not placed. It also cited bank account balances that did not belong to these people.

The most striking case is the same weather and tides chat. In the middle of an ordinary answer, Muse suddenly added "BTW, two bank updates" and listed two balances. Another employee used the agent on her personal phone. Muse reminded her to reply to her sister on Signal, a popular messenger with end-to-end encryption. She wrote in the post that this looked like a serious privacy violation. By her account, she had certainly not given the agent access to other apps and does not recall being asked for it.

Two questions need to be separated here. The first: whose data it was, that is, whether it leaked between different people. The second: what the agent has access to on the device in the first place. The publication makes clear that employees saw information in their chats that did not belong to them. What exactly broke inside Muse is not reported. Meta has not yet responded to TechCrunch's request for comment, so all conclusions about the causes remain speculation.

Why it matters

Muse is one of the fastest-growing apps in recent history: according to TechCrunch, it has more than 10 million downloads. While the agent was a toy for enthusiasts, failures like this would have remained curiosities. Now messages, orders, and financial information of millions of people pass through it. Any error in separating context turns into a leak of personal data.

Agents differ from ordinary chatbots in that they have memory and access to external services: email, messengers, bank notifications. The more sources are connected, the higher the cost of a mistake. A regular chatbot that malfunctions simply answers off-topic. An agent with access to several sources can bring someone else's letter or someone else's balance into the conversation.

User consent is a separate issue. The employee's statement that she does not recall being asked to grant access to other apps points to a problem with the interface. Even if permission was technically obtained, people must understand what the agent reads and when.

What agent builders should do

Several practical lessons follow from this story. These are general engineering rules, not an analysis of Muse's internals.

  • Separate user memory and context at the storage level, not just in the prompt. Do not rely on the model itself to "not show other people's data."
  • Verify the source of every fact before the agent inserts it into a response. Data should be bound to a specific account, and tool requests should carry its identifier.
  • Test for cross-account leaks deliberately: create several test profiles with unique markers in the data and check whether those markers surface for other users.
  • Request access to sensitive sources explicitly and in plain language, and do not surface financial data without a direct user request.
  • Prepare an incident response process in advance: logs, a fast way to disable the feature, and user notification.

Meta will likely have to explain publicly how these failures happened and how many ordinary users they affected. So far, employee messages only show that the company encountered them on internal devices. For the entire industry, the signal is simple: the faster an agent grows, the more careless data isolation costs.

Story

Key Questions Answered

What happened?

Reuters reviewed internal messages from Meta employees. They complained that Muse mixes conversations from different accounts and shows other people's personal data. According to them, the agent sent messages about requests to find a babysitter that no one had made. It gave career advice nobody had requested and reported orders for goods that they had not placed. It also cited bank account balances that did not belong to these people. The most striking case is the same weather and tides chat.

Why it matters?

Muse is one of the fastest-growing apps in recent history: according to TechCrunch, it has more than 10 million downloads. While the agent was a toy for enthusiasts, failures like this would have remained curiosities. Now messages, orders, and financial information of millions of people pass through it. Any error in separating context turns into a leak of personal data. Agents differ from ordinary chatbots in that they have memory and access to external services: email, messengers, bank notifications.

What agent builders should do?

Several practical lessons follow from this story. These are general engineering rules, not an analysis of Muse's internals. Separate user memory and context at the storage level, not just in the prompt. Do not rely on the model itself to "not show other people's data." Verify the source of every fact before the agent inserts it into a response. Data should be bound to a specific account, and tool requests should carry its identifier.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…