Incident sheet · /security

Filed
Read
3 min · 542 words
Severity
Routine

Meta's Muse AI agent leaks bank balances and cross-account chats

Meta employees report Muse leaked bank balances to the cent and read Signal messages without permission, per Reuters. The agent has topped 10 million downloads since September.

Meta’s AI agent Muse is mixing up conversations and exposing private data, report says
FeedMeta’s AI agent Muse is mixing up conversations and exposing private data, report says — tranmautritam / pixabay

Event log

  1. Reuters reviewed internal Meta posts showing Muse mixing up conversations across accounts and exposing private data.
  2. Muse sent an unprompted message citing a Capital One balance of $2722.58 and a Charles Schwab balance of $5899.57.
  3. Muse reminded an employee about a Signal message from her sister without granted access to other apps; she called it "an extreme privacy violation."
  4. Muse launched in September and has more than 10 million downloads, making it one of the fastest-growing apps on record.
  5. Meta said employees test different Muse versions than the public and that feedback drives safety and privacy improvements.

Meta's internal AI agent Muse has been mixing up conversations across accounts and exposing private user data, according to a Reuters report that reviewed internal company posts. The agent, launched in September and downloaded more than 10 million times, is one of the fastest-growing apps on record. That growth now shares a sentence with Capital One balances.

Employees reported that Muse surfaced content that did not belong to them. The list reads like a compliance officer's nightmare: childcare requests they never made, career advice they never asked for, product orders they never placed, and bank account balances belonging to other people.

What did the leaks look like?

In one instance, an employee told colleagues that Muse sent an extra, unprompted message into a chat about weather and tide information. The message read: "BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57."

To be clear about the mechanics: a financial disclosure, with named institutions and figures to the cent, injected into a conversation about tides. Nobody had asked for it.

A second case is arguably worse. An employee using Muse on her personal phone received a reminder to respond to a message from her sister on Signal — the end-to-end encrypted messaging app. She had not granted Muse 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," the employee wrote in her post, according to the report.

The implication is straightforward. If an agent can read notifications or messages from an encrypted app without a granted permission, the boundary between the assistant and the rest of the device is not a boundary. It is a suggestion.

What does Meta say?

A Meta spokesperson framed the incidents as part 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 to TechCrunch.

The spokesperson added a distinction that matters for deployment decisions: "Importantly, employees have access to different versions of Muse than the public, so they're often reviewing features, functionality, and behavior not available elsewhere."

That is a fair technical caveat. It is also a limited one. The versions employees test and the versions the public downloads ship from the same vendor, under the same architecture, with the same apparent appetite for data the user never handed over. Whether the public build exhibits the same cross-account behaviour is precisely the question Meta's statement does not answer.

What does this change for enterprise buyers?

For CIOs weighing AI agents with device-level access, the episode is a due-diligence data point, not a footnote. The failure mode here is not a jailbreak or a prompt injection. It is the agent volunteering third-party financial data and reading other applications without consent — the two failure modes access-governance reviews exist to catch.

The practical takeaway: demand evidence of permission scoping and data isolation before granting any agent broad device access, whatever the download numbers say. Ten million installs are a traction metric, not a security attestation.

via techcrunch.com (Source)

Filed under

  • ai-agents
  • data-privacy
  • access-governance
  • data-isolation
  • enterprise-ai
Share this article:

More from Tom Whitfield

Tom Whitfield

Show full bio

Market editor covering people, launches and day-to-day practice in enterprise information technology and the leaders who run at CIO Monitor.

188 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous articleNext article ›

Meta Muse AI Agent Leaks Bank Balances and Chats — CIO Monitor