Jolly, the mascot of Meta’s Muse AI agent
Meta employees found Muse, the company’s AI agent, reading encrypted messages and making up purchases. It also sent them chats that seemed meant for other people. That is according to more than two dozen internal posts reviewed by Reuters’ Katie Paul, published on Wednesday.
Staff raised concerns that Muse could be mixing up conversations between accounts and exposing private data, the posts showed. Muse has passed 10 million downloads since its September launch. Meta expects it to generate billions of dollars, Reuters reported.
Bank balances and a Signal reminder
In one post, an employee said Muse added an unprompted line to a chat about fishing weather.
“BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57,” the message read.
Another employee said Muse reminded her to reply to her sister on Signal, the encrypted messaging app.
“This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this,” she wrote.
A colleague found that Muse had permission by default to read push notifications on Android phones. One employee said Muse booked a non-refundable Airbnb stay without permission. Another said it offered to message a seller about a coffee table in what was “not my conversation”.
Reuters could not establish whether Muse was showing users other people’s data or making it up. In early November, a member of Meta’s Superintelligence team gave the internal feedback group an explanation. Built-in prompts push the agent to engage users between sessions, and it appeared to hallucinate messages in response.
“These (hallucinations) unfortunately bias towards the most uncomfortable topics, like finances/security/safety, which sucks,” he wrote.
How Muse reached Signal
Reuters worked with Signal staff and the AI Now Institute to repeat the Signal access. On an Android phone, Muse read messages through its notification permission. On a Mac, it used Apple’s accessibility and screenshot tools. The same requests failed on an iPhone.
Apple said its accessibility tools were built for people with disabilities. An Apple update this week closed that route on Macs, Reuters reported. It also added a new set of permissions for agents. In October, Apple tightened Full Disk Access because of AI agents. A journalist had said Muse read the messages on his Mac. Meta denied that claim at the time.
Meta’s response
Meta spokesperson Daniel Roberts called the examples “isolated incidents” and said Muse reaches other apps only through permissions users explicitly grant, Reuters reported. Many of the suspected account mix-ups were hallucinations that Meta had already fixed, he said. He declined to say how many there were.
Meta gave TechCrunch a further statement on Thursday.
“Importantly, employees have access to different versions of Muse than the public, so they’re often reviewing features, functionality, and behavior not available elsewhere,” a Meta spokesperson said.
Signal president Meredith Whittaker told Reuters that deploying agents with such wide access to people’s devices was “technologically reckless”. Two days later, Signal said it would block screenshots to keep AI agents out of its chats.
Get the TNW newsletter
Get the most important tech news in your inbox each week.