Meta Muse Told an Employee Someone Else's Bank Balance
Meta's Muse agent has been mixing up conversations across accounts in employee testing, according to a Reuters investigation built on internal company posts (https://www.reuters.com/investigations/meta-bets-big-muse-ai-agent-early-safety-mishaps-surface-2026-11-18/, dated November 18). TechCrunch picked it up on November 19 at 17:38 UTC and added Meta's comment a few hours later: https://techcrunch.com/2026/11/19/metas-ai-agent-muse-is-mixing-up-conversations-and-exposing-private-data-report-says/
The examples are the kind no privacy policy can talk its way out of. Employees said Muse surfaced childcare requests they never made, career advice they never asked for and product orders they never placed. One got an unprompted add-on in a chat about weather and tides: "BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57." Those balances did not belong to that employee. Another, using Muse on a personal phone, got a reminder to reply to a sister's message on Signal and wrote that they never granted it access to other apps or remember being asked.
Meta's answer: employee testing exists to catch exactly this, and employees run different versions of Muse than the public, with features that are not available elsewhere. That is a fair point about process and a weak one about architecture. A test build that can pull one person's bank balance into another person's chat means the data from different accounts was reachable from the same place to begin with. Muse launched in September and has passed 10 million downloads, so the public build had better not share that plumbing.
Two failures are tangled together here and they need different fixes. Cross-account leakage is an isolation bug, the oldest kind of bug in multi-tenant software, and agents make it worse because they read widely and write confidently. The Signal reminder is a consent bug: the agent acted on an app the user did not knowingly connect. Every personal-agent launch this fall has bragged about how much context it can pull in. Muse is the first big one to show the bill for that, and the fix is boring and old: separate tenants, explicit grants, and logs a user can read.
← Back to all articles
The examples are the kind no privacy policy can talk its way out of. Employees said Muse surfaced childcare requests they never made, career advice they never asked for and product orders they never placed. One got an unprompted add-on in a chat about weather and tides: "BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57." Those balances did not belong to that employee. Another, using Muse on a personal phone, got a reminder to reply to a sister's message on Signal and wrote that they never granted it access to other apps or remember being asked.
Meta's answer: employee testing exists to catch exactly this, and employees run different versions of Muse than the public, with features that are not available elsewhere. That is a fair point about process and a weak one about architecture. A test build that can pull one person's bank balance into another person's chat means the data from different accounts was reachable from the same place to begin with. Muse launched in September and has passed 10 million downloads, so the public build had better not share that plumbing.
Two failures are tangled together here and they need different fixes. Cross-account leakage is an isolation bug, the oldest kind of bug in multi-tenant software, and agents make it worse because they read widely and write confidently. The Signal reminder is a consent bug: the agent acted on an app the user did not knowingly connect. Every personal-agent launch this fall has bragged about how much context it can pull in. Muse is the first big one to show the bill for that, and the fix is boring and old: separate tenants, explicit grants, and logs a user can read.
Comments