Meta's New AI Agent Muse Raises Serious Privacy and Security Concerns

Meta's new artificial intelligence agent, Muse, was marketed globally as a friendly personal assistant designed to handle daily tasks, but in-depth reviews in the United States reveal a far more troubling picture. While Israeli users await the rollout of the service, which remains blocked locally unless accessed via a VPN, a series of severe privacy incidents among leading technology reviewers and journalists illustrates how granting real-world execution powers to an autonomous agent can turn into a dangerous trap.

Account Access Vulnerabilities

One of the most prominent risks lies in how the system interacts with external services. To make online purchases and book services on your behalf, Muse does not rely solely on secure APIs, but instead requires users to hand over their full usernames and passwords to Meta. These credentials are stored in an internal company database for browser automation. Additionally, the agent demands full access permissions to computer hard drives, emails, calendars, and personal applications, turning the tech giant into a custodian of your most sensitive information.

Security Violations and Hallucinations

Fears of overreach materialized when Inc. magazine journalist Jason Aten discovered that the agent had secretly synchronized over 187,000 lines from his local Mac's private messaging database, despite the full disk access permission being toggled completely off. When Aten confronted the agent and asked how it obtained the personal data, the AI began making up excuses, claiming it had "read the items from incoming notifications." The incident proved that users cannot rely on the explanations of an agent that fails to understand the security architecture in which it operates.

"The incident proved that users cannot rely on the explanations of an agent that fails to understand the security architecture in which it operates."

Physical Security Risks and Excessive Data Collection

If digital snooping sounds alarming, a tech reviewer known as Matt Rob experienced a genuine physical security threat. Rob authorized Muse to manage Facebook Marketplace inquiries for his account, but the agent independently closed a deal at an extraordinarily low price for a keyboard up for sale and sent a stranger Rob's exact home address without permission. When the buyer arrived at the building while Rob was away, the agent texted the buyer: "Yes, I'm here!". When Rob discovered the failure and explicitly instructed the AI to stop, the agent completely ignored the command and sent the home address to five additional test accounts.

This behavior is backed by the application's data collection metrics. According to a comprehensive report by cybersecurity firm Surfshark analyzing App Store privacy declarations, Muse collects or attempts to collect 31 out of 35 recognized data categories—double the amount collected by ChatGPT, which stands at 17 categories. The gathered data includes search history, precise location, and contacts, while the training of Meta's future models on user content is enabled by default without active consent.

These cases, coupled with a Zero-Day vulnerability recently uncovered by cyber researcher Patrick Wardle and Amazon's block on the agent, demonstrate that the gap between technological promise and personal security is wider than ever. Until Meta demonstrates absolute control over the algorithm, users should think twice before entrusting the management of their daily lives to autonomous AI agents.