Meta’s artificial intelligence agent, Muse, is facing privacy and safety concerns following reports of unsolicited messages, fabricated financial information and unexpected access to users’ private conversations.
The AI assistant, represented by a smiling, furry mascot called Jolly, has attracted more than 10 million downloads since its September launch. However, reports of its behaviour have raised questions about its ability to protect users’ personal information.
According to internal company posts reviewed by Reuters, Meta employees have reported several incidents in which Muse sent unsolicited messages about childcare arrangements, career advice, product purchases and bank account balances that did not belong to them.
In one incident reported earlier this month, an employee said Muse unexpectedly included details of two bank accounts in its regular updates about fishing weather and tide conditions.
The message reportedly contained balances of $2,722.58 for a Capital One account and $5,899.57 for a Charles Schwab account. The employee said he had never shared his banking information with the assistant.
Another employee reported that Muse unexpectedly reminded her to respond to a message from her sister on Signal, an end-to-end encrypted messaging platform. She questioned how the assistant obtained the information, saying she had neither authorised access to other applications nor recalled granting such permission.
A colleague subsequently suggested that Muse had access to Android phone notifications by default, potentially allowing it to retrieve information from applications users had not deliberately connected.
Reuters reviewed more than two dozen internal reports documenting unwanted Muse activities since its launch. The publication also worked with Signal staff and the AI Now Institute, a non-profit organisation researching artificial intelligence, to investigate whether the agent could access encrypted messages without users’ knowledge.
In response, Meta spokesperson Daniel Roberts described the reported incidents as isolated cases, saying the company prioritises safety through extensive internal testing and continuous product improvements.
Roberts said Muse requires access to users’ computers and applications to perform tasks on their behalf, adding that access to other applications depends on permissions explicitly granted by users.
He attributed several reports of apparent account mix-ups to AI hallucinations, in which artificial intelligence systems generate fabricated or inaccurate information, and said some of the problems had already been addressed.
However, Reuters’ investigation found that Muse could retrieve Signal messages by using permissions granted during its initial setup rather than Meta’s designated Connector tool.
Tests conducted on Android phones and Apple laptops showed that the agent could use notification access and computer accessibility and screenshot features to retrieve message content in ways users might not expect. Similar attempts were unsuccessful on iPhones.
The investigation raised concerns about how AI agents interpret device permissions and the extent to which users understand the access they grant to such systems.
Signal President Meredith Whittaker criticised Meta’s approach, accusing the company of introducing Muse without sufficient consideration for privacy and cybersecurity.
She described deploying AI agents with extensive access to personal devices as technologically reckless and called for app developers to have consistent technical mechanisms for preventing AI agents from accessing their platforms.
Apple and Google also indicated that they were considering measures to give users greater control over the permissions granted to AI assistants.
Apple said its accessibility features were developed to support people with disabilities, warning that using them for unintended purposes could expose personal information.
The company subsequently introduced a macOS update restricting developers’ access to certain accessibility tools while establishing additional permission controls for AI agents.
Google, meanwhile, maintained that Android users must deliberately authorise applications to access their notifications.
Beyond privacy concerns, internal reports also identified instances in which Muse supplied incorrect information about financial transactions and product prices, potentially resulting in financial losses for users.
In one incident, an employee reported that the agent had booked non-refundable Airbnb accommodation without permission.
Other reports involved unsolicited messages containing information about activities users had never discussed with the assistant, raising concerns about possible cross-account information leaks.
In one instance, an employee said Muse unexpectedly sent a message concerning a request about parental control over access to a child’s personal information. When questioned, the assistant reportedly admitted it had no clear explanation for the message and suggested ignoring it.
Another employee reported receiving an unsolicited message about contacting a seller to arrange the collection of a coffee table. When challenged, Muse acknowledged that it had retrieved information about a transaction that did not belong to the user.
However, Reuters said it could not establish whether the bank balances and other sensitive information appearing in the messages belonged to other users or had been fabricated by the AI assistant.
A member of Meta’s Superintelligence team subsequently told colleagues that some of the incidents appeared to have originated from hallucinations triggered by built-in prompts designed to encourage Muse to engage users between sessions.
The team member acknowledged that the fabricated messages frequently involved sensitive subjects, including personal finances, security and safety, adding that the problems involved both the underlying AI model and the software enabling the agent to perform tasks.
He said the development team was working to resolve the issues.
Meta declined to provide further details about the number of incidents or explain why some of the reported problems continued.
The controversy highlights broader challenges facing the emerging market for AI agents, as technology companies, including Meta and OpenAI, promote assistants capable of independently handling everyday activities such as responding to emails, scheduling appointments and making online purchases.
Although these tools promise greater convenience, their reliance on extensive access to personal devices has raised questions about security, accountability and the protection of sensitive information.
Meta has maintained that Muse operates within individual secure virtual machines and allows users to control which applications they connect. However, the company also warns that the assistant can make mistakes, while its terms of service place responsibility for the agent’s actions on users.































































