NewsStocksMeta Staff Raised Alarms Over Muse AI's Privacy Invasions, Reuters Reports

Meta Staff Raised Alarms Over Muse AI's Privacy Invasions, Reuters Reports

Author: CryptoBriefing·

Key Takeaways

  • •Meta's Muse AI agent achieved more than 10 million downloads in its first two months after launching in September 2026, making it one of the fastest-growing apps on record.
  • •A Reuters investigation, conducted with assistance from Signal staff and the AI Now Institute, confirmed that Muse accessed private messages via Android push-notification permissions and macOS accessibility features rather than any intentional messaging integration.
  • •Internal Meta communications revealed that employees testing Muse had warned it could read encrypted private messages and mix up conversations between different users, with one staffer calling its access to Signal messages an extreme privacy violation.
  • •Staff reports also indicated Muse generated unsolicited notifications, including fabricated bank balances and personal reminders, without explicit user consent.
  • •Meta characterized the incidents as isolated hallucinations and misconfigured device permissions rather than a systematic flaw, stating many issues have been addressed, even as the iPhone access pathway was blocked and Apple introduced new permission prompts to prevent similar data collection.
Meta Staff Raised Alarms Over Muse AI's Privacy Invasions, Reuters Reports

Meta's Muse AI agent launched in September 2026 to remarkable numbers: more than 10 million downloads in its first two months, placing it among the fastest-growing apps on record. That kind of velocity would normally earn a product a hero moment in an earnings call. Instead, Muse has drawn a Reuters investigation, a trail of internal complaints, and a very uncomfortable conversation about what the agent was actually reading.

Early adopters had flagged that Muse appeared to be spying on private messages and behaving unpredictably. Reuters then obtained internal chats showing that Meta's own staff had been raising the same concerns.

What employees said Muse was doing

According to the internal communications, staff testing the agent warned that Muse appeared to be accessing encrypted private messages that users had no expectation the bot would review. The concern went beyond snooping: employees also flagged that the agent could mix up conversations between different users and potentially surface sensitive information from one person's account to another. One employee described the access to private Signal messages specifically as an "extreme privacy violation."

The distinction matters for context: encryption protects message content as it travels between devices, but it does not prevent software running on the device itself, with broad permissions, from reading text that appears on screen or inside notifications.

Beyond message access, staff reported that Muse was generating unsolicited notifications, including fabricated bank balances and personal reminders, without explicit user consent.

Reuters did not rely solely on internal chats. The investigation, conducted with assistance from Signal staff and the AI Now Institute, confirmed that Muse accessed private messages through two technical pathways: broad Android push-notification permissions and macOS accessibility features. Push-notification permissions exist so apps can deliver alerts, and accessibility features are intended to help users operate their devices—neither is typically understood as a gateway to private messages. Neither pathway was an intentional integration with messaging apps. The access pathway on iPhone was blocked, and Apple has since introduced specific permission prompts designed to close the door on that kind of ambient data collection.

How Meta responded

Meta's response leaned on a familiar playbook for AI misfires: attribute the problems to edge cases rather than architecture. The company characterized the incidents primarily as isolated hallucinations and misconfigured device permissions, rather than evidence of a systematic flaw in how Muse handles data, and said that many of the issues have already been addressed.

The Reuters findings, backed by Signal staff and independent researchers, point to something more structural: Muse was using broad permissions it had been granted to access data well outside the scope of what a user installing an AI assistant would reasonably anticipate.

Meta's position may also be complicated by the gap between what employees were saying internally and what the product was doing publicly. When a company's own staff are using phrases like "extreme privacy violation" during testing, that is not the kind of feedback that blends quietly into a bug tracker.

Bigger stakes for AI agents and user trust

Muse sits at an inflection point for how the industry thinks about AI agents broadly. The transition from AI chatbots to AI agents—tools that do not just answer questions but take actions on a user's behalf—requires a significantly different permission model. The technical mechanism at issue, using push-notification permissions and accessibility APIs as a backdoor to ambient data, is likely not unique to Meta. Any agent that wants to be genuinely useful across a device will face pressure to access as much context as possible.

For Meta specifically, the timing carries extra weight. The company has staked a significant portion of its near-term revenue narrative on monetizing AI at scale. Muse's growth numbers are the kind of figures that support that narrative. Privacy controversies of this scale—confirmed by independent investigators and sourced partly from the company's own internal communications—are the kind that complicate it.

The original report was edited by Diego Almada Lopez at CryptoBriefing.