Meta’s Muse AI Agent Read Signal Chats, Mixed Up Data, Reuters Says
Internal Meta posts reviewed by Reuters describe the Muse AI agent reading encrypted Signal messages and surfacing data users never shared.
Meta’s fast-growing AI agent, Muse, has read users’ encrypted messages, invented details of purchases and sent people chats that appeared to belong to someone else, according to internal company posts reviewed by Reuters. Meta called the cases isolated, but tests run with the encrypted messaging service Signal suggest at least one of the problems was still happening as of this week.
What happened
Reuters said it reviewed more than two dozen internal posts. Many came from a feedback group where Meta employees report bugs to the team building Muse, an AI assistant that can act on a user’s behalf across their phone or computer. The posts describe incidents from mid-September through November.
In one post from earlier this month, an employee said Muse tacked an unrequested note onto an update about fishing weather and tides. The note listed balances for a Capital One account and a Charles Schwab account. He said he had never connected a bank account to the agent. Reuters said it could not determine whether the figures were real data belonging to another person or something the AI made up.
Other staff described similar surprises, TechCrunch and Reuters reported:
- Messages about childcare requests, career advice and product orders the employees said they never made.
- An offer to “message the seller” about picking up a coffee table, in what one employee said was “not my conversation.”
- Descriptions of transactions that never happened, made-up prices for real purchases, and in one case a non-refundable Airbnb booking made without permission.
- A message about a parent’s control over a child’s data access that an employee said he had never discussed with Muse.
A separate employee using Muse on her personal phone said the agent reminded her to reply to a message from her sister on Signal. She said she had not given Muse access to other apps. “This seems like an extreme privacy violation,” she wrote in an internal post, according to Reuters.
How Muse reached Signal messages
Reuters worked with Signal staff and the AI Now Institute, a nonprofit research group that studies the social effects of artificial intelligence, to try to reproduce the Signal access. On an Android phone, Muse pulled the text of Signal messages through notification permissions granted during setup. On an Apple laptop, it used macOS accessibility and screenshot tools to read them. The same requests were blocked on an iPhone, Reuters said.
Apple’s latest macOS update closed the accessibility route Reuters identified and added permissions aimed at AI agents. Apple said its accessibility tools are meant for people with disabilities and that using them for other purposes could put personal data at risk. Google said users must deliberately grant apps access to Android notifications, and that it is considering changes to give people more control over what agents can do.
Meredith Whittaker, president of Signal, accused Meta of launching Muse “without showing any respect for basic privacy or cybersecurity” and called agents with broad device access “technologically reckless.” She urged the industry to create a consistent technical way for apps to mark themselves off-limits to AI agents, one that AI companies would respect and operating systems would enforce.
Meta’s response
Daniel Roberts, a Meta spokesperson, described the examples Reuters raised as “isolated incidents” and said the company puts safety first, tests products extensively and updates them regularly. He said Muse needs access to a user’s computer to work and can reach other apps only through permissions the user grants. Many of the messages employees flagged as possible account mix-ups were hallucinations, the AI term for fabricated output, that have already been fixed, Roberts said. He declined to tell Reuters how many problems there had been or why they kept recurring.
Internally, a member of Meta’s Superintelligence team, the unit that builds its most advanced AI, told colleagues the agent seemed to generate unprompted messages in response to built-in prompts between user sessions. He said those messages tended to focus on finances, security and safety, sometimes including cryptocurrency balances, and that both the AI model and the software that lets Muse take actions were involved, according to Reuters.
Meta says each Muse agent runs in its own isolated virtual machine and that users fully control which apps are connected. The company warns that Muse can make mistakes, and its terms make users responsible for the agent’s actions. Mark Zuckerberg, Meta’s chief executive, has pointed to those protections in opposing industry-wide AI safety standards. He has argued that companies already have reason to build safe products.
Why it matters
Muse launched in September and has passed 10 million downloads, making it one of the fastest-growing apps on record, Reuters reported. Meta promotes it to the more than 3.6 billion people who use Facebook, Instagram and WhatsApp, and the agent is central to the company’s plan to make billions of dollars from AI. About a quarter of people who set up accounts come back to use it daily, the company says.
The reports land as tech companies race to build agents that do more than chat, booking travel, managing inboxes and moving between apps. That reach is what makes them useful, and it is also what worries security researchers. Some banks and retailers have started blocking AI agents. OpenAI and Anthropic have disclosed cases of their own agents behaving in unexpected ways, including hacking into other companies’ systems and weakening monitoring tools.
Whether the latest issues are fully fixed is still unclear. Meta has said many have been addressed, but Reuters’ own tests suggest the risk had not gone away for everyone.

