One Meta employee was asking Muse about weather and tides for a fishing trip when the AI agent added something nobody asked for. “BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57.”

The balances were not his. He had never shared a bank account with Muse and never talked about banking with it. “I’m afraid it’s sending other’s bank info to me,” he wrote in an internal post, according to Reuters.

That post is one of more than two dozen internal messages Reuters reviewed in which Meta staff flagged Muse mixing up conversations between accounts and surfacing private information that belonged to someone else. The reports land two months after Muse launched in September and passed 10 million downloads, making it one of the fastest-growing apps on record.

What Meta employees reported about Muse

The internal posts describe a pattern, not a single glitch. Employees said Muse sent them messages about childcare requests they never made, career advice they never asked for, and product orders they never placed. Bank balances showed up uninvited in ordinary chats.

Another employee got a different kind of surprise on her personal phone. Muse reminded her to reply to a message from her sister on Signal, the end-to-end encrypted app. She had not connected Signal to Muse, and she did not recall giving it access to other apps.

“This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this,” she wrote, according to the report.

Reuters did not stop at the posts. It worked with staff at Signal and the AI Now Institute, a nonprofit that studies AI’s impact on society, to test whether Muse could reach messages in other apps. The tests indicated the problem was ongoing and could be affecting people outside Meta too.

How Muse reached messages in other apps

The tests reproduced the behavior on Android phones and Apple Mac computers, while similar attempts on iPhones were blocked. That split points to device permissions, not a break in Signal’s encryption, as the way in.

On Android, a colleague found Muse had been set up by default with permission to access push notifications, which let it read message previews from apps the user never connected. On a Mac, Muse used accessibility and screenshot permissions, features built to help people with disabilities use their computers, to reach message content.

Nobody has shown Muse cracked Signal’s encryption. The reporting shows something narrower: the agent used ordinary setup permissions to read things users never meant to share.

One open question cuts both ways. Reuters could not confirm whether the bank balances Muse displayed were real data belonging to another person or numbers the model made up. Hallucinated figures would still be a serious failure for an agent trusted with personal tasks, but they are a different failure from leaking a stranger’s account.

What Meta says

Meta spokesperson Daniel Roberts described the examples as “isolated incidents,” saying the company prioritizes safety, conducts extensive internal testing and keeps updating its products.

A separate Meta spokesperson told TechCrunch that employee testing is the point of the process. “Employee testing is core to the product development process and the entire point is to get feedback, so we can implement safety and privacy protections and improve products,” the spokesperson said in an email.

Meta also said some of the incidents involved AI hallucinations and have been fixed. That may cover the bank balances. It sits less comfortably next to the Signal test, where Reuters found the behavior still happening after the reports surfaced.

Why the Meta Muse AI agent privacy problem matters

Muse is not a side project. Meta built it as the consumer face of its AI push, an agent meant to live inside messaging, handle errands, and remember what matters to each user. That job only works if the agent keeps one person’s life separate from everyone else’s.

Mixing up conversations breaks the core promise. A personal agent that cannot tell whose bank balance it is looking at, or which messages belong to whom, is not ready to sit between people and their money, their families, or their private chats.

The timing makes it harder for Meta to wave away. Regulators in the US and Europe are already watching how AI agents handle personal data, and a Reuters investigation built on Meta’s own internal posts is the kind of evidence that draws questions from lawmakers.

Users can limit the damage today. On Android, check whether Muse has notification access and switch it off if you never meant to grant it. On a Mac, review its accessibility and screen-recording permissions.

Meta has not said how many people outside its staff have seen the problem. Until it does, the employees’ own read stands: an agent built to know everything about you should not be showing you someone else’s life.

Frequently asked questions

What did Meta’s Muse AI agent do wrong?

According to a Reuters review of internal Meta posts, Muse mixed up conversations between accounts and showed employees childcare requests, career advice, product orders and bank balances that were not theirs.

Did Muse really show other people’s bank balances?

One employee reported Muse adding Capital One and Charles Schwab balances to a fishing weather chat. Reuters could not confirm whether the figures were real data from another person or AI hallucinations.

How did Muse access Signal messages?

Tests by Reuters with Signal and the AI Now Institute reproduced the behavior on Android and Mac. On Android, Muse read message previews through push notification permissions; on Mac, it used accessibility and screenshot permissions. It did not break Signal’s encryption.

What has Meta said about the Muse privacy reports?

Meta called the cases “isolated incidents” and said some involved AI hallucinations. A spokesperson said employee testing exists to surface problems so the company can add safety and privacy protections.

Sources: Reuters, TechCrunch, Rappler