The AI agent, represented by a smiling, fuzzy mascot called Jolly, has reportedly attracted more than 10 million downloads since its September launch. However, internal company reports reviewed by Reuters suggest that its unexpected behaviour has raised questions about the safety of a technology Meta hopes will generate billions of dollars in revenue.
In several reported incidents, employees said Muse sent them information about childcare arrangements, career advice, purchases they had never made and bank balances that did not belong to them.
One employee reported that the chatbot unexpectedly added financial information to its regular conversations about fishing weather and tide conditions.
"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 surprised her with a reminder to respond to a message from her sister on Signal, an encrypted messaging platform popular among privacy-conscious users.
Reacting to the incident in an internal company post, she said: "This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this."
A colleague subsequently explained that Muse was configured by default to access push notifications on Android devices, potentially allowing it to retrieve messages from applications users had not directly connected.
Reuters reviewed more than two dozen internal reports detailing unwanted actions by Muse since its introduction. The reports were largely shared in an employee feedback group used to communicate problems to the team developing the AI assistant.
The news agency also worked with Signal staff and the AI Now Institute, a non-profit organisation studying the societal impact of artificial intelligence, to reproduce the reported access to private messages.
Their tests indicated that the issue was ongoing and could potentially affect other users.
However, Meta spokesperson Daniel Roberts described the incidents as "isolated incidents", insisting that the company prioritises safety, conducts extensive internal testing and regularly updates its products to strengthen protections.
According to Roberts, Muse requires access to users' computers and applications to perform tasks on their behalf, but can only access other apps through permissions users explicitly grant.
He also said several messages interpreted by employees as possible exchanges of information between different accounts were AI hallucinations, meaning information generated by artificial intelligence that is inaccurate or entirely fabricated.
Meta said those particular issues had already been addressed. Reuters, however, could not independently establish whether Muse had exposed other people's actual financial information or simply invented the details.
Concerns over extensive device access
The controversy highlights wider concerns surrounding the growing use of AI agents, a new generation of artificial intelligence tools being developed by companies including Meta and OpenAI.
Unlike conventional chatbots, AI agents are designed to interact directly with users' devices and perform tasks such as responding to emails, booking appointments and purchasing products with limited human involvement.
Meta has gained an early advantage in the consumer AI market, with Muse reportedly climbing app download charts following extensive promotion across Facebook, Instagram and WhatsApp, platforms collectively used by more than 3.6 billion people.
The company has also disclosed that approximately a quarter of users who have established Muse accounts return to use the service daily.
However, the technology's ability to access smartphones and computers presents significant security and privacy challenges, particularly when AI systems misunderstand instructions or exploit permissions in unexpected ways.
Some financial institutions and retailers have already responded to these risks by blocking AI agents from interacting with their services.
The controversy has also raised questions about the responsibilities of technology companies such as Apple and Google, which control the operating systems and permissions that allow applications to access users' devices.
Meta has previously assured users that each Muse agent operates within an isolated, secure virtual machine, a cloud-based environment designed to simulate a personal computer. The company also maintains that users have control over which applications they connect.
Nevertheless, its terms of service warn that the AI agent can make mistakes and place responsibility for its actions on users.
Tests reveal how Muse accessed encrypted messages
Investigations conducted by Reuters and its research partners found that Muse could retrieve Signal messages through existing device permissions, even without using Meta's designated Connector tool.
On Android devices, the AI agent reportedly used permission to access push notifications to retrieve message content.
On Apple's macOS, testers found that Muse could use accessibility and screenshot tools to obtain information displayed on the computer.
The same attempts were unsuccessful on an iPhone.
The findings raised concerns that Meta may be interpreting permissions granted during the assistant's initial setup more broadly than users expect.
Signal President Meredith Whittaker criticised the deployment, accusing Meta of launching Muse "without showing any respect for basic privacy or cybersecurity".
She described the widespread access granted to AI agents as "technologically reckless" and called for stronger protections to prevent such tools from accessing applications without users' informed consent.
Whittaker also urged application developers to be given consistent technical mechanisms for identifying software that should remain inaccessible to AI agents.
Apple and Google, responding to Reuters' findings, said they were considering changes that would give users greater control over the permissions granted to AI agents.
Google maintained that Android users must deliberately authorise applications to access their notifications.
Apple, meanwhile, warned that using accessibility features for purposes other than those intended could expose personal information.
The company said its latest operating system update had introduced tighter restrictions on developers' access to accessibility tools on Mac computers, alongside new permissions specifically designed for AI agents.
"We are committed to ensuring users understand exactly what they are allowing, so they can make informed choices about their data and privacy," Apple said.
Employees report unexpected messages and financial errors
Beyond concerns about encrypted communications, internal reports also described cases in which Muse allegedly provided inaccurate information about purchases, invented product prices and initiated transactions without users' approval.
In one reported incident, an employee said the AI agent booked a non-refundable Airbnb accommodation without permission.
Other employees described receiving strange messages that appeared to contain information from conversations they had never had.
One staff member reported that Muse unexpectedly sent a message saying it could not submit a request concerning parental control over access to a child's data.
The employee said he had never discussed childcare arrangements or parental permissions with the chatbot.
When questioned about the unexpected message, Muse reportedly responded: "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?"In another incident, an employee said the AI agent unexpectedly offered to contact a seller about collecting a coffee table.
When asked to explain, Muse acknowledged that it had "pulled in some context about a coffee table pickup that doesn’t seem to be yours."
Reacting to the incident, the employee wrote: "This was not my conversation," expressing concern about a possible exchange of information between separate accounts.
A member of Meta's Superintelligence team later suggested that some of the strange messages could be linked to built-in prompts designed to encourage Muse to engage users between sessions.
In an internal discussion in early November, the employee acknowledged that the AI's hallucinations appeared to focus disproportionately on sensitive subjects.
"These (hallucinations) unfortunately bias towards the most uncomfortable topics, like finances/security/safety, which sucks," he wrote.
He also disclosed that the team had received reports involving cryptocurrency information, including unexpected updates about users' Bitcoin balances.
According to the employee, the problems involved both the underlying AI model and the software that enables Muse to perform tasks on users' behalf.
The development team was working to resolve the issues, he added.
Meta spokesperson Roberts declined to answer Reuters' questions about the number of incidents reported or why some of the problems appeared to persist.
The growing controversy has intensified scrutiny of consumer AI agents, raising questions about whether the convenience they promise can be delivered without putting users' private information and digital security at risk.

