Meta’s Muse artificial intelligence agent, introduced in September, accessed private messages, reported purchases that had not occurred and displayed financial information whose authenticity could not be established, according to an investigative report published by Reuters on Wednesday.
Muse has been downloaded more than 10 million times since launch. Meta expects the technology to generate billions of dollars in revenue over the longer term, but its expansion has been accompanied by security and privacy concerns. Reuters examined more than two dozen internal company posts in which employees described unexpected behavior by the agent.
Private messages accessed through device permissions
One Meta employee said Muse unexpectedly reminded them of a private message received through the Signal messaging app, even though they said they had not granted the assistant access to it. Reuters repeated the test with experts from Signal and the AI Now Institute and was able to reproduce the problem.
On Android, the agent read message contents through access granted to notifications. On Apple’s macOS, it used accessibility and screenshot functions; on iPhones, testers were able to prevent the same access. The system did not break Signal’s end-to-end encryption, but instead used permissions previously granted on the device in ways users might not have anticipated.
Questionable financial data and purchases
Another Meta employee wrote that, during an ordinary conversation about fishing, Muse unexpectedly displayed the balances of two bank accounts: $2,722.58 for a Capital One account and $5,899.57 for a Charles Schwab account. The employee said they had not shared banking information with the assistant and feared the figures might belong to other users.
Other accounts described Muse reporting purchases that had not happened or giving incorrect prices for items that had actually been bought, causing financial losses. One employee said the agent booked a non-refundable Airbnb accommodation without permission.
Reuters could not establish whether the financial information that appeared to belong to other users was real or had been invented by the agent. Meta said the incidents were isolated and that many of the suspected data leaks were caused by AI hallucinations—false information generated by the system. The company said it was continuing to work on fixing security problems.
Meredith Whittaker, president of Signal, called it technologically irresponsible to deploy AI agents that can gain broad access to users’ devices. Both Apple and Google said they were examining ways to tighten user permissions. Apple also released an operating system update this week that largely closed the macOS access route identified by Reuters.



