NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

An Agent That Saw Someone Else's Bank Balance

Three items came back in one output: messages, a bank balance, an order history. Two were about money.

2026-11-21 · 4836 words · NeuroAI
An Agent That Saw Someone Else's Bank Balance

Three items came back in one output: messages, a bank balance, an order history. Two were about money. None belonged to the person reading the screen.

The output came from Muse, Meta's internal agent, according to a TechCrunch report on November 20. Employees said it surfaced other users' messages, balances, and orders.

Meta said the outputs were not real data.

Three things that are not one thing

A message is a communication record. A balance is account state. An order is a transaction trail. Each has an owner, read conditions, and usually its own screen.

One query returned all three. The boundary did not hold.

A boundary failing is not a wrong answer. A wrong answer has the wrong content. A leak has the wrong object. The first is fixed by editing a sentence. The second by changing who the system can reach.

Seeing your own order is a feature. Seeing someone else's balance is a permission crossing a line. On a keyboard the two are identical. Inside the system they are a layer apart.

One sentence, two questions

Meta said the outputs were not real data. That answers a question about content.

Employees raised a question about placement: should this be on this screen at all?

A person looking at someone else's balance in their own window, even a fabricated one, is looking at something that should not be there. Real or fake is one layer. Boundary is another. The first is arguable. The second was designed, not computed, and designed things can be inspected. Computed things can only be sampled.

Citi moved a clause

Citi updated its terms: customers are responsible for activity from account access through third-party AI tools, even if hacked.

"Even if hacked" takes theft out of the exemptions.

The clause governs the entrance. Money leaves by a route, and liability follows it. The route used to be online banking, the mobile app, the branch. Now there is a fourth, installed by the customer. The first three were opened by the bank. The fourth was not. When the entrance changes hands, liability moves with it.

The bank is not grading the tool. It is writing one line: you opened this road, you carry what happens on it.

The account boundary is not at the bank's door. It is at the moment the customer grants access.

A model behind a list

OpenAI restricted GPT-6 Cyber to authorized security researchers.

The more a model can do, the fewer people can touch it. The restriction sits at the model end, and the mechanism is a roster. A roster is the cheapest access control and the coarsest. It stops people. It does not stop interfaces.

Three companies, three positions. Meta at the data end, OpenAI at the model end, Citi at the contract end. None coordinated, and all three drew the reachable perimeter first. Those are the three doors an agent passes: where the data comes from, what the model may do, who pays when it goes wrong. The third is skipped most often, because it is not in the product documentation.

Aggregates at one end, a person at the other

ChatGPT opened a beta of Businesses. Companies can claim a page, update information, connect a catalog, and see aggregated data on where they were mentioned, clicked, and discovered. The platform reports about 1.2 billion weekly active users. Some compliant US merchants can deploy labeled Business Agents.

1.2 billion is a platform number. What reaches one company is three words: mentioned, clicked, discovered.

A merchant learns 1,000 people clicked, not which.

In the same week, one product erases the individual and one incident exposes him. The erasure is the business. The exposure is the accident.

The scale is whether you can see a particular person. That is not a gap in technical difficulty. It is a gap in permission design. Aggregation is the language of product; the individual is the language of audit. Both run in one system, and only permissions keep them apart. When they fail, the merchant sees a person, not 1,000.

Three verbs

Apple is hiring engineers for proactive agents. The posting lists three verbs: write code, control apps, operate devices.

Three verbs, three levels of permission. Code touches files. App control touches interfaces. Device operation touches the machine. Files roll back. Interfaces reload. A device operated on your behalf does not, because something happened in the physical world.

Permission tiers follow this order: read, write, act. The line is not in the model. It is in the interface.

The third is hardest to undo. An agent that misreads can be corrected. An agent that presses the wrong button may have moved the money.

Job postings describe capabilities, not boundaries. Capability first, boundary later is the order every agent team follows: the reverse makes an ugly page, and ugly pages do not attract the people who want to build this. The boundary comes afterward, usually after an incident.

Permission before capability

Four events from one week: a leak, a clause, a restriction, a job posting.

They converge. What an agent can do depends on what it can reach. Capability is the result. Access is the premise.

Capability can be demonstrated. Permission can only be set in advance.

An agent that reads email and one that reads email and can start a transfer differ by one capability level and several levels of consequence. The step between them is not called smarter. It is one more interface. Capability is continuous. Interfaces are discrete.

Interfaces are designed, not trained. Training teaches a model where to click. Design decides where it may click. Learning belongs to the model. Allowing belongs to the product.

The third verb operates the device, which means the device can be out of human hands.

What separates them is permission.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "An Agent That Saw Someone Else's Bank Balance," 21 November 2026, https://neuroai.site/a/na-corp-meta-muse-agent-1121

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Companies & Stack” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories