Meta staff raise privacy concerns over Muse as Reuters tests its access to private messages
Tests found routes into private messages on Android and macOS. Whether unexpected financial outputs contained anyone’s real data remains unresolved.
Listen to this story
The audio brief
Story brief
3 key pointsReuters’ testing suggests Muse’s privacy boundary depends not only on Meta’s Connector but also on device-level permissions, creating a route to message content that app developers may not control. Employees’ reports of conversations and bank balances that seemed to belong to others do not establish an actual cross-account leak: Reuters could not verify whether the details were retrieved or hallucinated. The episode puts pressure on Meta’s safeguards and on operating-system makers to provide enforceable ways for apps to block AI agents.
- 01
On Android, Muse could read Signal messages using permissions granted at setup; macOS access used accessibility and screenshot features, while iPhone requests were blocked.
- 02
Muse has recorded more than 10 million downloads since its September launch.
- 03
Employees described unexpected context, including bank names and precise balances during a fishing conversation; Reuters could not confirm the information’s source.
Meta employees testing its Muse AI agent received unsolicited messages about requests they had never made and financial information they did not recognize, according to a Reuters investigation. Their internal warnings raised fears of private data crossing between accounts. Testing by Reuters also found that Muse could retrieve private messages through device permissions outside Meta’s dedicated connection tool.
Muse has recorded more than 10 million downloads since its September launch, Reuters reported.
Device permissions opened another route
Reuters worked with Signal staff and the AI Now Institute to test message access. As detailed by Ars Technica, the results differed across operating systems:
- Android: Muse could read Signal messages using permissions granted during initial setup, rather than going through Meta’s Connector tool for connecting services.
- iPhone: Requests to retrieve messages were blocked in the testing.
- macOS: Muse retrieved messages using accessibility-related tools—features intended to help people with disabilities operate their computers—including screenshot tools.
Messages that seemed meant for someone else
Muse offered one employee help messaging a seller about a coffee table, then said it had pulled in context that did not appear to belong to them. The employee flagged the exchange as “not my conversation” and raised concern about a possible leak between accounts.
In another example, a conversation about fishing produced an unsolicited update naming two banks and giving precise account balances.
Meta points to prompts and a fixed bug
Meta spokesperson Daniel Roberts called the employee reports isolated incidents and said many apparent account mix-ups were hallucinations—fabricated AI outputs—that the company had addressed.
Meta’s David Singleton had also described an instruction-delivery failure. An update telling Muse to seek confirmation before certain actions instead affected every message. Singleton characterized the resulting behavior as a product bug causing hallucinations, not a security or privacy issue, and said it had been fixed.
A spokesperson told Ars that employees use different versions of Muse and often test features unavailable to the public.
The message-access findings prompted a different demand from Signal President Meredith Whittaker: give app developers consistent technical ways to mark their services off-limits to AI agents. She argued that operating systems should enforce those boundaries.
Sources
- arstechnica.comMuse privacy invasions are freaking out Meta staffers, internal chats show
- japantimes.co.jpAs Meta bets big on Muse AI agent, early safety mishaps surface
Reader comments
Newest comments first. Replies stay oldest first.