Meta’s artificial intelligence agent Muse is facing serious privacy questions after employees reported that it displayed unfamiliar conversations, unexpected bank account details and information from private messaging apps. According to a Reuters investigation published on November 18, staff raised concerns that Muse could be mixing up conversations between different accounts and exposing sensitive information belonging to other people.
The problems emerged during internal testing of the agent, which has attracted more than 10 million downloads since launching in September 2026. Meta is positioning Muse as a central part of its AI business, a helper that can automate activities and complete tasks across a user’s digital services. The reports from inside the company suggest that the very capability that makes the agent useful, its ability to reach across apps and services on a person’s behalf, is also what makes it risky.
AI Agent Displays Unexpected Bank Details
Reuters reviewed internal reports in which Meta employees described responses about activities they had never discussed with Muse, including childcare requests, career advice and product orders that had nothing to do with them, alongside bank balances they did not recognise. In one incident, an employee asking about weather and tide conditions unexpectedly received a message containing two bank balances: a Capital One balance of 2,722.58 dollars and a Charles Schwab balance of 5,899.57 dollars. The employee said they had never shared banking information with Muse and feared the agent might be showing another person’s financial details.
Reuters could not establish whether the balances belonged to real accounts or were fabricated by the AI system, and that uncertainty sits at the heart of the story. A made-up bank balance is a very different problem from a genuine leak of someone’s financial data. But from the employee’s point of view, the distinction barely mattered: they had asked about the weather and been shown numbers that looked exactly like a stranger’s bank accounts, with no way of knowing which it was.
Private Signal Messages Raise Concerns
In another incident, a Meta employee using Muse on her personal smartphone received an unexpected reminder to respond to a message from her sister on Signal, an end-to-end encrypted messaging app designed to keep private conversations from outside eyes. The employee said she had not knowingly authorised Muse to access other applications and questioned how it obtained information about her messages.
Reuters also tested the issue with Signal and the AI Now Institute and reproduced instances of Muse accessing Signal message content. The investigation found that certain device permissions, including Android notification access and macOS accessibility settings, could allow the agent to read information from other applications without using Meta’s dedicated app connection system. Similar testing on an iPhone did not reproduce the access, suggesting the problem depends on how much freedom the operating system gives an app to look at what other apps are doing.
Meta Responds to Privacy Allegations
Meta said employee testing plays an important role in identifying problems before products become widely available. A company spokesperson told TechCrunch that employees sometimes test versions of Muse with features not available to the public, implying some of the reported behaviour may come from experimental builds rather than the version ordinary users download. Meta also described some incidents as isolated and said certain unexpected financial information may have been generated by the AI rather than obtained from real accounts.
The company maintained that users must grant relevant device permissions and can control which applications are connected. Meta said it is working to address issues involving both the AI model and the software that carries out tasks, a recognition that the problem spans the agent’s judgement about what to show and the plumbing that lets it reach into other apps.
More Than 10 Million Downloads
Muse has become one of the fastest-growing AI applications, passing 10 million downloads since its September launch. Meta is betting heavily on it, presenting the agent as the next step beyond the chatbot: not just something you talk to, but something that acts for you across your services. That vision only works if users trust the agent with deep access to their digital lives, and the Reuters findings strike directly at that trust.
Signal criticised how the technology was deployed and called for stronger privacy protections, while Apple and Google have also examined ways to improve user control over sensitive permissions. Apple has already restricted one of the macOS access methods identified during testing.
Although the reported incidents do not establish that Muse has exposed genuine financial information belonging to other users, they highlight the privacy risks of AI agents that can interact with personal messages, applications and accounts. The tension is fundamental. An agent that cannot see your messages, your calendar and your accounts cannot do very much for you. An agent that can see all of it can, in principle, show it to the wrong person, or invent something that looks just like it. Until companies can prove which one is happening, users have good reason to check exactly what permissions they have granted and which apps they have connected.
