Meta’s Muse AI Builds Hourly Profiles on Users’ Social Circles

Researchers have uncovered that Meta’s new AI agent, Muse, maintains detailed and regularly updated files on individuals within users' social networks. Nearly a month after its launch, investigations into the tool’s operations revealed that it openly shares internal instructions when queried by users.

Muse’s Hourly Data Collection Process

A particularly unsettling aspect of the system is an hourly process directed at creating and updating “a page for every person in the user’s life,” as reported by Wired. This mechanism allows the AI to track relationships even with people who do not use Muse themselves.

Karan Joshi, an independent AI safety researcher, extracted a trove of these instructions and system prompts via the app’s chat interface, sharing his findings with Wired. Separately, The Future Society published system instructions derived from an exported copy of a Muse virtual machine.

Researcher Concerns Over Intimacy

Joshi told Wired that the collected data suggests Meta aims to understand users' real-world relationships deeply. “They’re trying to know you like a friend, which is honestly pretty creepy,” he stated regarding the system skills and prompts fed into the agent.

These revelations follow Meta’s September 8 launch of Muse, positioned as a personal AI agent capable of handling tasks such as sending emails, booking travel, shopping, and tracking goals. To perform these functions, users grant Muse access to a wide range of their apps and accounts. Since its release, the application has reached the number-one spot on Apple’s U.S. App Store chart for free iPhone apps.

Security Flaws and Corporate Pushback

Privacy and security issues surfaced almost immediately after the launch. In September, Amazon blocked Muse from accessing its online shop, stating that Meta had not provided advance notification of the agent’s intent to use the platform. Amazon also raised concerns about how Muse manages customer credentials and account data.

Additionally, Meta recently patched a zero-day vulnerability in the Mac version of Muse. This flaw could have allowed malware already present on a user’s computer to hijack the agent and exploit previously granted permissions.

System Instructions Detail Relationship Tracking

According to system instructions published by The Future Society, the relationship pages generated by Muse can include a person’s name, their connection to the user, recurring topics like upcoming moves or savings goals, and significant dates such as birthdays and anniversaries.

The agent also maintains a history of each relationship, recording events, milestones, closeness levels, interaction patterns, and what the relationship appears to need at any given moment.

Meta Defends Contextual Necessity

A Meta spokesperson defended the data collection in an emailed statement to Gizmodo, arguing that context is essential for the agent’s utility. “For any agent to be useful and actually help you achieve your goals, it needs to have context about you and those you interact with,” the spokesperson said.

The company stated that Muse gathers this information based on public data and what users choose to share. This method allows the AI to remember, for example, that an invoice sender is the plumber hired for bathroom work, or which flowers a spouse prefers. Meta added that this information is stored in each user’s dedicated secure virtual machine, which they claim cannot be accessed by other users’ agents.