Meta's AI agent Muse is handing employees bank balances that aren't theirs
Employees say the internal agent surfaced messages, orders, and financial data belonging to other people, a boundary failure that turns an AI rollout into a governance test.

Meta employees say the company's AI agent, Muse, surfaced messages about childcare requests they never made, career advice they never asked for, product orders they never placed, and bank account balances that did not belong to them, per a TechCrunch report. For executives running enterprise AI rollouts, the incident reframes the risk from model quality to permission architecture, where a single retrieval error can move real personal data into the wrong person's session.
Meta employees say the company's AI agent, Muse, has been handing them other people's information. According to a TechCrunch report, workers described Muse surfacing messages about childcare requests they never made, career advice they never asked for, product orders they never placed, and bank account balances that did not belong to them. Read that list again, because the categories matter. These are not creative hallucinations about the weather. They are fragments of real, personal, identifiable activity, arriving in the wrong person's feed. For anyone running an enterprise AI rollout, that is the difference between an embarrassing demo and a reportable incident.
The distinction is the whole story. A language model that invents a plausible-sounding bank balance is a quality problem. A system that retrieves an actual balance belonging to someone else is a data-boundary failure, and data-boundary failures are the ones that end up in front of regulators, lawyers, and works councils. The four examples employees gave span the full spectrum of sensitivity: family logistics, career conversations, purchase history, and financial data. Childcare requests and career advice tend to live in private messages. Order history and account balances are the kind of records that privacy regimes treat as protected. If a single agent is pulling from all of them at once, the question is not whether the model is smart. It is what the model was allowed to read.
That question is now the central one for every company racing to put agents on top of internal data. The pitch for enterprise agents is simple and genuinely compelling: instead of searching five systems, you ask one assistant and it goes and finds the answer. To do that, the agent typically needs broad read access across mail, chat, documents, ticketing, and HR systems. Historically, that access was scoped per application, and a human moved information between them. Agents collapse those walls by design. The upside is speed. The downside is that a single permissioning mistake, a stale index, or a retrieval bug can route one employee's records into another employee's session, at machine speed, without anyone noticing.
Regulators have been circling this exact failure mode. Europe's GDPR requires data to be processed for specified purposes and kept to the minimum necessary, and it gives employees rights over their own workplace data. The EU AI Act adds obligations around transparency and risk management for higher-risk uses. In the US, the FTC has spent years scrutinizing how companies handle sensitive data, and a growing patchwork of state privacy laws imposes similar duties. None of that requires a breach in the classic sense. Mishandled internal data can be enough.
For Meta, the timing is awkward. The company has staked a large part of its identity on AI, from its open-weight Llama models to the assistant it has pushed across its apps. Internal tooling is where that ambition meets its own workforce, and employees are a uniquely vocal constituency. They have legal rights, they talk to each other, and in Europe they often have formal representation. A tool that leaks a colleague's bank balance into your chat window does not need to be widespread to become a trust problem. It only needs to happen once to someone who tells everyone.
The second-order effects are where this gets expensive. Enterprise AI programs live or die on adoption, and adoption depends on employees believing the tool is safe to use. Once that belief cracks, you get shadow bans: people stop asking the agent anything sensitive, which quietly destroys the productivity case you sold to the board. You also get legal exposure. If personal data belonging to employees or customers is surfacing in the wrong context, that can trigger notification duties, regulatory inquiries, and in some jurisdictions consultation with worker representatives. And you get a security question: an agent with broad read access is a high-value target, and its logs are a map of everything it can see.
The practical response is not to abandon agents. It is to treat them like any other system with privileged access. That means classifying data before the agent touches it, scoping permissions to the narrowest set that still makes the product useful, logging every retrieval so you can reconstruct what a user saw and when, and red-teaming the retrieval layer specifically for cross-user bleed. It means deciding in advance what the kill switch looks like and who can pull it. And it means asking vendors, internal or external, a blunt question: show me how you prevent one user's data from appearing in another user's answer.
The strategic stakes are straightforward. Every large company is now under pressure to show AI progress, and internal agents are the easiest place to demonstrate it. That makes them the easiest place to ship something under-governed. The companies that win the next two years will not be the ones that deployed the most agents fastest. They will be the ones whose agents can be trusted with the data they were given, because the alternative is a headline that costs more than the tool ever saved. Meta's Muse is a reminder that the hard part of AI was never the model. It was the permissions.
This story's Key Insights and Take-aways are locked.
Create a free account to unlock Executive Actions for one credit.
Register to UnlockAlways free for Executives Club members. Join the Club
More in Technology
Anthropic mathematician claims Claude helped crack the Riemann hypothesis
If the proof holds, AI has toppled math's most famous open problem; if it does not, it is a case study in announcing breakthroughs before anyone verifies them.
AI's $4 Trillion Wipeout Erases Nasdaq's Gains for the Year
A brutal week took roughly $4 trillion off the AI sector and left the Nasdaq flat for the year, and the damage ran the full length of the supply chain.
Anthropic mathematician says AI proved the Riemann hypothesis. Verification will take time.
A single unverified claim from inside Anthropic now sits against more than 160 years of open mathematics, and the gap between asserting a proof and surviving peer review is the entire story.


