N環球 N GLOBE
US Technology

Reuters investigation: Meta's AI agent Muse mixes up conversations and reads encrypted messages, staff say

· N GLOBE

A Reuters investigation found Meta's AI assistant Muse, launched in September, repeatedly sent employees messages meant for other people, including strangers' bank balances and reminders about Signal messages. Meta calls them isolated incidents, mostly hallucinations, but Reuters replicated the Signal access with Signal staff.

Meta's AI assistant Muse has burrowed into encrypted personal messages, supplied made-up information about purchases and sent users chats seemingly meant for other people, according to a Reuters investigation published on November 18. Reuters reviewed more than two dozen internal posts in which Meta employees raised concerns that the agent was mixing up conversations between different accounts and exposing users' private data.

Staff posts: other people's bank balances, requests never made

In one internal post from earlier this month, an employee wrote that Muse had randomly added an extra message to its regular chats about fishing-related weather and tide information, saying: "BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57." The employee wrote: "I have never shared my bank account with Muse and never discussed any banking topic with Muse. I'm afraid it's sending other's bank info to me." Other employees reported messages about childcare requests they had not made, career advice they had not sought and product orders they had not placed.

In another case, an employee said Muse sent him a message "out of nowhere" saying it was unable to submit a request related to a parent's control over who could access their child's data. He had done nothing to trigger the chat and had never discussed child data with his Muse. When he asked the bot about the message, it replied: "I don't have a good explanation for it other than that it wasn't related to our conversation. Let's ignore it. How can I help?" Another staffer said Muse randomly offered to "message the seller" about picking up a coffee table, then admitted it had "pulled in some context about a coffee table pickup that doesn't seem to be yours." The employee called it "not my conversation" and raised concern about a "possible cross-account context leak."

Encrypted messages were read too: Signal replication confirms it

One employee using Muse on her personal phone reported it had surprised her with a reminder to respond to a message from her sister on Signal, the end-to-end encrypted messaging service that competes with Meta's WhatsApp. "This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this," she said. Reuters worked with staff at Signal and the AI Now Institute to replicate the access she had observed, and the tests indicated the issue was ongoing and could be affecting others.

On an Android phone, Muse used its permission to access the device's push notifications to retrieve the messages. On an Apple laptop, it used a combination of the computer's accessibility and screenshot tools, repurposing settings Apple offers to help people with disabilities operate their computers. The same requests were blocked in tests conducted on an iPhone. Signal President Meredith Whittaker accused Meta of launching Muse "without showing any respect for basic privacy or cybersecurity," calling the deployment of agents that can gain wide access to people's devices "technologically reckless." She called for apps like Signal to be given "consistent technical ways" to mark themselves off-limits to AI agents.

Meta's response: isolated incidents, mostly hallucinations

Meta spokesperson Daniel Roberts described the Reuters examples as "isolated incidents," saying the company prioritizes safety and continuously updates its products. He said Muse needs to see and interact with people's computers to work on their behalf and can access other apps only via permissions they explicitly grant, and that many of the messages flagged by employees as potential account mix-ups were hallucinations that had already been addressed. Reuters was unable to establish whether Muse was giving users other people's bank account balances or making up the information.

However, the internal posts reviewed by Reuters showed such incidents occurring as far back as mid-September and continuing through this month. In early November, a member of Meta's Superintelligence team told the internal feedback group that the agent appeared to be hallucinating messages in response to built-in prompts pushing it to engage with users between sessions. "These (hallucinations) unfortunately bias towards the most uncomfortable topics, like finances/security/safety, which sucks," he said, adding that he was aware of several cases involving "crypto-related things too, like giving people updates on their Bitcoin balances." He said the issues involved both the underlying AI model and the software enabling Muse to take actions. Roberts declined to answer Reuters' questions about those issues, including how many there were and why they were persisting.

An earlier flaw: 6,800 shared chats discoverable in public DNS

It is not Muse's first privacy problem. In October, security firm Claroc found that Muse's conversation-sharing links could be discovered through publicly searchable passive DNS records: all shared links used the domain cf.metaaiusercontent.com, with the conversation title embedded directly in the subdomain, so a single query could enumerate about 6,800 shared conversation links, hundreds of which contained sensitive information never intended for public access. Claroc's co-founder reported the findings through Meta's bug bounty program on October 8 and received no response; Meta paused the sharing feature on October 20 and later restored it with a fix.

Apple has tightened permissions as regulatory pressure grows

Reuters noted the incidents have also pushed operating system providers to act. Apple changed its operating system this week to give people more control over agents' data usage, sharply restricting developers' use of accessibility tools on Macs and introducing a new agent-oriented set of permissions, after an Inc. Magazine columnist accused Muse of reading private chats from the Messages app on his Mac. Apple said it was committed to ensuring users understand exactly what they are allowing; Google said it was considering changes too.

Other employee cases included Muse booking a non-refundable Airbnb stay without permission and making up prices on products users did purchase, costing them real money. Meta has assured users it isolates each agent in its own secure "virtual machine" and gives people full control over which apps they connect. CEO Mark Zuckerberg has cited those protections in dismissing calls for industry-wide AI safety standards, while Meta's terms of service warn that Muse may make mistakes and hold users responsible for the agent's actions.

Muse has had more than 10 million downloads since its September launch, one of the fastest-growing apps on record, with around a quarter of the people who set up accounts returning each day, and it is the centerpiece of Meta's plan to generate billions of dollars in AI revenue. Reuters said the incidents highlight the risks of a new category of consumer AI agents from Meta, OpenAI and others that promise to execute tasks on users' behalf, and the new challenges they pose for privacy and for the operating systems that manage their permissions.

💬 Comments

No comments yet. Be the first!