SDSignal Desk

Meta’s own testers say Muse dropped bank balances that weren’t theirs into their chats

· Updated · Signal Desk newsroom

Analysis of reporting by TechCrunch. How we work

Image: TechCrunch

A Reuters report on internal Meta posts describes Muse surfacing other people’s messages, orders and account balances. Meta calls it the point of employee testing; the episode deepens Muse’s trust problem.

Key takeaways

  • Per Reuters via TechCrunch, Meta employees reported Muse surfacing childcare requests, career advice, orders and bank balances that weren’t theirs.
  • One employee said Muse appended two bank balances, at Capital One and Charles Schwab, to an unrelated weather-and-tides chat.
  • Another said Muse reminded her to reply to her sister on Signal, despite not recalling granting access to other apps.
  • Meta says employee builds differ from the public version and that catching such issues is the purpose of testing.
  • Muse launched in September and has more than 10 million downloads, with a string of earlier privacy disputes.

The worst kind of agent bug

Assistants get things wrong all the time. What Reuters describes is a different category of failure: information crossing between accounts. According to internal posts Reuters reviewed, Meta employees testing Muse saw messages about childcare requests they never made, career advice they never asked for, orders they never placed, and bank balances that weren’t theirs. In one case, an otherwise mundane chat about weather and tides ended with an unprompted note listing a Capital One balance and a Charles Schwab balance. The employee didn’t hold those accounts.

If accurate, that is the nightmare scenario for any agent that is allowed to read finances and messages: someone else’s private data arriving in your conversation, which also means your data may be arriving in someone else’s.

Meta’s answer, and its limits

Meta told TechCrunch that employee testing exists to surface exactly these problems, and that staff run different versions of Muse than the public does, often with features not yet released. That is a reasonable explanation for some of the reports, and nothing in TechCrunch’s summary shows the cross-account mix-ups reaching public users. It is not a full answer, though. Meta didn’t say which behaviors were limited to internal builds, whether the leaked balances belonged to real people, or what has been fixed.

The second example is harder to wave off. An employee using Muse on her personal phone got a reminder to respond to her sister on Signal, the end-to-end encrypted app, and wrote that she did not recall granting access to other apps. Whether that came from notification access, a permission prompt she forgot, or something else isn’t clear from the report. But it rhymes with incidents we have covered before.

A pattern, not an incident

Muse’s short life has included a September dispute over whether it read a columnist’s private messages, an Apple change to macOS Full Disk Access prompted partly by that case, and a Threads scare in which the agent invented an explanation for how it saw Messages. One hands-on review found it better at surveilling than helping, nudging users toward linking bank accounts and scanning inboxes.

Each episode had a plausible innocent explanation. Taken together, they describe an agent whose access outruns its user’s understanding of that access, and sometimes Meta’s own. Muse is one of the fastest-growing apps on record, past 10 million downloads, and Meta has made it the center of its personal-superintelligence pitch. That growth is exactly why the bar for isolation between accounts has to be absolute.

Who carries the risk

Muse users
Anyone who linked bank, messaging or shopping accounts should review permissions until Meta explains what was fixed.
Meta
Internal dissent leaking to Reuters adds to a run of trust problems just as Muse becomes core to its strategy.
Regulators
Cross-account data exposure, if it reached any public user, would be a far more serious matter than a hallucination.
Rival agent makers
A reminder that memory and account isolation, not model quality, may decide which consumer agents people trust.

What to watch

  1. Whether Meta confirms if any cross-account exposure affected public users, and whose balances appeared.
  2. Changes to how Muse requests and displays access to third-party apps like Signal.
  3. Regulatory inquiries in the US or EU citing the Reuters report.
  4. Further internal posts or whistleblower accounts about Muse’s memory and data handling.

FAQ

What did Meta employees report?

Per Reuters, Muse showed them messages, orders, advice and bank balances that didn’t belong to them, plus a Signal reply reminder one employee didn’t recall enabling.

Did this affect public users?

Unknown. Meta says employees use different versions of Muse than the public. The report doesn’t establish public exposure.

What did Meta say?

A spokesperson said employee testing is meant to surface feedback so Meta can add safety and privacy protections.

Is this Muse’s first privacy issue?

No. Earlier disputes involved Messages access on macOS, an invented notification explanation and a reported security flaw.

Read the original

Continue at the source.

TechCrunch

Companies: Meta

More on Meta

All Meta coverage →