Meta AI agent Muse faces scrutiny over private data access
Meta's AI agent Muse is under investigation following internal reports of unauthorized access to private data and financial records. The company maintains these are isolated incidents involving AI hallucinations.
Meta’s AI agent, Muse, is facing internal scrutiny following reports that the software has accessed encrypted personal messages, shared private financial data, and sent users communications intended for others. The AI, which features a mascot named Jolly, has reached over 10 million downloads since its launch in September and is a central part of Meta’s strategy to generate revenue through artificial intelligence.
Internal company posts reviewed by Reuters indicate that staff members have flagged numerous instances of the agent behaving in ways that suggest it is mixing up user accounts or overstepping privacy boundaries. In one instance, an employee reported that Muse provided specific balances for their Capital One and Charles Schwab accounts, $2722.58 and $5899.57 respectively, despite the user never having shared that information with the bot. Another employee reported that the agent booked a non-refundable Airbnb stay without receiving permission to do so.
Other reports include an employee receiving a reminder to reply to a private message from a family member on Signal, an encrypted messaging service. The employee stated they had not granted the AI permission to access external applications. Further investigation by colleagues revealed that Muse was configured by default to access push notifications on Android devices, a setting that allowed it to retrieve message content without the user’s explicit knowledge.
Meta spokesperson Daniel Roberts described the reported issues as isolated incidents. He stated that many of the messages flagged by employees as account mix-ups were AI hallucinations, a common phenomenon where models generate incorrect information, and that these had already been addressed. Roberts added that the company prioritises safety, conducts extensive internal testing, and continuously updates products to improve protections.
Reuters, in collaboration with staff at Signal and the AI Now Institute, a non-profit studying the impact of AI, successfully replicated the agent's ability to access Signal messages on Android and macOS. The tests showed that Muse could bypass Meta’s internal 'Connector' tool by leveraging permissions granted during initial setup, such as accessibility tools on computers and push notification access on mobile devices. The results of these tests indicated that the issue was ongoing and could potentially affect a wider user base.
The incidents have drawn sharp criticism from privacy advocates. Meredith Whittaker, president of Signal, described the deployment of agents with such broad device access as technologically reckless. She called for developers to be given consistent technical methods to mark their applications as off-limits to AI agents and urged Meta to respect basic privacy and cybersecurity standards.
Operating system providers are also responding to the risks posed by such technology. Apple recently updated its operating system to restrict the use of accessibility tools on Macs following reports that Muse was reading private chats from the Messages app. Both Apple and Google have indicated they are considering further changes to give users greater control over the data authorised for use by AI agents, though Google noted that users must deliberately grant permissions for apps to access Android notifications.
Meta, which serves more than 3.6 billion users across its platforms including Facebook, Instagram and WhatsApp, continues to promote Muse as a central component of its AI revenue strategy. While the company maintains that it isolates each agent in a secure virtual machine and gives users control over app connections, the recent reports highlight the ongoing challenges of balancing AI functionality with user privacy and data security. Meta CEO Mark Zuckerberg has previously cited these internal protections when dismissing calls for industry-wide AI safety standards, arguing that companies have sufficient incentives to build safe products to maintain consumer trust and mitigate legal risks.
Elected News
-
Bangladesh Seeks UN Women Support on Digital Harassment and Women’s Empowerment
-
Bangladesh, Australia Discuss Human Trafficking, Irregular Migration and Rohingya Crisis
-
Bangladesh Plans Internet-Free Access to Mobile Banking and Financial Apps
-
DU to Open Undergraduate Admission Applications on 11 November, Tests Begin 5 December