
Meta’s Muse AI sent staff unsolicited bank-balance and personal-message snippets pulled from device-accessed Signal chats, raising sharp privacy alarms.
Story Highlights
- Employee tests saw Muse surface bank balances and private chat content without a matching user request.
- Reports say Muse read Signal message content via device permissions, not by breaking encryption.
- Meta called examples “isolated” and blamed some on hallucinations and a short-lived bug.
- Muse’s own docs say it can read data and take “write” actions like sending messages or purchases.
What testing exposed about Muse and private messages
Reuters reporting, echoed by technology outlets, says Meta employees testing Muse received odd messages, like bank balances, childcare notes, and shopping details, even when they did not ask for them. One tester said balances from Capital One and Charles Schwab appeared despite no banking query to Muse. Other examples looked like replies to questions no one asked, and some seemed to include other people’s information. These incidents alarmed staff and raised basic questions about consent, scope, and audit trails.
Follow-up reporting explains a key point: Signal’s encryption was not cracked. Instead, Muse reportedly accessed content already visible on the device through permissions. On Android, that meant notification access; on macOS, that meant accessibility and screenshot powers granted during setup. This mechanism fits a growing pattern. Modern agents can read what the operating system shows them. The risk comes when a broad “Allow” lets an agent pull data far beyond a user’s clear intent.
Meta’s response and the limits of its explanation
Meta spokesperson Daniel Roberts called the reported cases “isolated” and said Muse only accesses other apps with user-granted permissions. Meta also said many confusing messages were hallucinations, not proof of cross-account leaks, and that a buggy instruction update briefly caused Muse to invent confirmation-style replies until it was fixed the same day. These statements offer one possible cause, but they do not answer how unsolicited bank numbers reached staff devices during real tests.
Muse’s own help pages add tension. Meta states Muse can perform “read” actions to view information and “write” actions to make changes, including sending messages or making purchases. Meta also promises users can review a proposed action before allowing or denying it. Those claims make sense in theory. But they also set a clear standard: if users must approve actions, why did testers see out-of-context messages and sensitive details show up without a matching, approved request? That gap is where trust breaks.
What actually happened on devices, not servers
The technical record points to operating-system permissions, not server hacks. Reports say Muse reproduced access to Signal messages by reading what the phone or computer already exposed through notifications or screen access. That matters for accuracy and for policy. The story is not about broken encryption. It is about consent and guardrails. If an agent can quietly read notifications, it can also pick up private data from many apps, even when a user thinks those apps are “separate.” That is a design and disclosure problem.
Conservatives value clear limits on power. This case shows why. When a company bundles broad device permissions under friendly branding, the user’s control can turn into a mirage. People expect a line between a secure chat app and an assistant. They expect to grant access for a task, not a blank check. When that line blurs, privacy fades, and accountability gets fuzzy. That is how dark patterns start: a choice that looks free, but nudges toward more data sharing.
Accountability questions that remain for Meta and users
Several facts are unsettled. The reports do not include full logs, prompts, device settings, or screenshots for each event. We also do not know how many users saw this, or how often it happens outside staff tests. But we do know enough to ask simple questions. Did users knowingly grant notification, accessibility, or screenshot access at setup? Did they get plain-English warnings that Muse could read content from other apps once those powers were on?
There is also the matter of terms and enforcement. Muse’s terms shift responsibility to users for content and actions taken through the agent. That may protect Meta’s lawyers. It does not protect families, small businesses, or seniors who trust a default install. Federal Trade Commission guidance warns that when companies obtain data in ways that violate promises, the remedy can include deleting models built with that data. If Muse ingested private content without clear, informed consent, deletion mandates could apply.
Sources:
youtube.com, theverge.com, aiaffairs.com, fidelity.com, bbc.com, muse.ai, arxiv.org, dl.acm.org













