Skip to content
LIVE
The Executives BriefThe Executives Brief

Meta's Muse told a staffer her $2,722.58 balance. Meta calls it a hallucination.

Internal posts seen by Reuters show Meta's own employees unnerved by an AI agent that surfaced Signal messages and bank figures, just as JPMorgan and Amazon move to wall agents out.

ByHessa Al-FalehBusiness Desk, The Executives Brief
·4 min read
Meta's Muse told a staffer her $2,722.58 balance. Meta calls it a hallucination.
Executive summary

Meta employees flagged its Muse AI agent after it volunteered private-seeming details, including a staffer's Capital One and Charles Schwab balances and a reminder about a sister's Signal message, which Meta says were hallucinations and opt-in notification access. The episode lands as JPMorgan CEO Jamie Dimon and Amazon refuse to let agents touch customer accounts or shop their platforms, turning permission design into the next competitive battleground in AI.

Meta's Muse AI agent told one of the company's own employees her bank balances. Not because she asked it to check them, and not while she was doing anything related to money. According to internal posts seen by Reuters, the employee was chatting with Muse about fishing conditions, weather and tides when the agent volunteered: 'BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57.' Meta later confirmed the message was a hallucination, not a real read of her accounts. That distinction matters enormously, and it is also exactly why the episode rattled the people building the product.

The bank balances were not the only message that landed badly. A Meta staffer reported that Muse sent her a reminder on her personal phone to respond to a message from her sister on Signal, the encrypted messaging service. 'This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this,' she posted. A colleague said he received unsolicited chats about apps he had never connected, and discovered Muse was set up by default with permission to access push notifications on Android phones. Other messages raised eyebrows too: a question about parental control over access to a child's data, and a conversation about picking up a coffee table that triggered concern over a 'possible cross-account context leak.'

Meta's answer is that this is what employee testing looks like. A member of the company's Superintelligence team said the agent appeared to be hallucinating because of built-in prompts designed to engage users between sessions. 'These (hallucinations) unfortunately bias towards the most uncomfortable topics, like finances/security/safety, which sucks,' he said, adding that other cases involved 'crypto-related things too, like giving people updates on their Bitcoin balances.' A Meta spokesperson said employee testing is core to product development and is meant to generate feedback so privacy safeguards can be implemented, noting that employees have access to different versions of Muse than the public does and that some functions are not available elsewhere. On the Signal messages, the spokesperson said Android users can grant apps notification access, an opt-in that is not specific to Muse. For Mac, Meta said, letting Muse work on your behalf requires clicking allow for accessibility and screen recording, then confirming again on a settings screen.

Scale is what turns a quirky internal thread into a board-level issue. Muse launched in September and has been downloaded more than 10 million times, making it one of the fastest-growing apps ever. It is also central to Meta's plan to generate billions of dollars in revenue from AI. An agent that hallucinates about your bank balance during a fishing chat is a funny screenshot inside a company that can fix it. The same agent, at 10 million downloads and climbing, is a trust problem in public, where the screenshot travels faster than the patch.

The walls are already going up. JPMorgan Chase CEO Jamie Dimon said the bank will not allow agents to access customer accounts and act on their behalf. Amazon will not let Muse shop on its e-commerce platform, and is already fighting to keep out Perplexity's AI agents. The logic is straightforward: an agent that acts on your behalf needs credentials, context and permission to touch systems built to be touched only by a human who logged in. Every company that says no is making a judgment about liability, not about technology. For Meta, that means the addressable market for Muse may be defined less by what the model can do than by which counterparties will let it in the door.

Raising those barriers has a cost. The network upgrades involved may be contributing to a sudden surge in outages. In recent weeks, Chase, Wells Fargo, United Airlines, Apple and Microsoft all saw temporary service disruptions. U.S. officials have said there is no evidence of cyberattacks, suggesting software patches were deployed before all the bugs were worked out. Sen. Mark Warner, vice chair of the Senate Intelligence Committee, told the Financial Times that AI tools can find vulnerabilities faster than companies can safely fix them, and that the pressure to fix them right away is straining systems never designed for that speed. Mark Giles, director of Downdetector parent company Ookla Research, told Fortune's Catherina Gioino there is nothing linking the outages. 'This is a big blip, but it's kind of on par for the course for what we're seeing in terms of those longer-term trends,' he said. 'Most of the time they're fairly self-contained and remedied quite quickly. It's the ones where you can't withdraw money or you can't access your emails that really get the headlines.'

For executives, the takeaway is not that AI agents are dangerous. It is that permissioning is now a product feature and hallucination is now a compliance risk. The Muse complaints were not about a model being wrong on a trivia question; they were about an agent appearing to know things it should not know, on a personal phone, about family messages and account balances. That is the exact category of failure that regulators, bank risk committees and enterprise security teams are already primed to scrutinize, and it arrives while the same companies are patching systems fast enough to cause visible outages. The companies that win the agent era will be the ones that can show, in plain language, what an agent can see, when it was granted that access, and how a user revokes it. Meta has 10 million downloads and a public explanation. Its peers have a preview of the questions their own boards will ask next.

Executive ActionsLocked

This story's Key Insights and Take-aways are locked.

Create a free account to unlock Executive Actions for one credit.

Register to Unlock

Always free for Executives Club members. Join the Club

More in Business