Meta's Muse AI Is Compiling Hourly Dossiers on Users and Their Unregistered Contacts

Millions of people who never downloaded Meta's new artificial intelligence agent are now the subjects of hourly surveillance dossiers. The company's Muse assistant, launched in September 2026, actively compiles detailed files on the friends, family, and colleagues of its users. This automated profiling sweeps up unregistered contacts without their consent or knowledge.[1][6]

The data collection operates continuously in the background of the popular application. Every 60 minutes, Muse updates its internal pages on any individual mentioned across at least three core communication channels: connected emails, text messages, and chat histories. The system maps out the user's entire social network, tracking exactly who matters to them and why.[1][2]

This capability represents a significant shift from passive social media algorithms to active relationship monitoring. Muse is explicitly instructed to infer user goals that have not been stated out loud. It achieves this by analyzing the context of private conversations and cross-referencing them with publicly available web data.[1][6]

The depth of the profiling extends far beyond basic contact information. Internal system instructions direct the agent to record relationship milestones, shared interests, and the underlying dynamics of social groups. Muse actively logs both "tensions and alliances" among a user's acquaintances to better predict their behavior.[1][2]

Extracting the system prompts

The extent of this background profiling was uncovered in early October 2026 by independent AI safety researcher Karan Joshi. By interacting with the application's standard chat interface, Joshi successfully extracted over 50 pages of Muse's underlying system prompts. These internal instructions revealed the hourly dossier updates and the mandate to map human relationships.[2]

"They're trying to know you like a friend, which is honestly pretty creepy," Joshi noted after reviewing the extracted data. The prompts show that Meta wants the agent to understand the nuances of real-world connections. This requires the software to constantly analyze the emotional weight of everyday digital communications.[2]

The agent uses this relationship history to tailor its interactions and suggest real-world actions. It notes inside jokes, memorable phrasing, and the specific needs of a relationship at any given moment. If a user's friend recently moved or celebrated a birthday, Muse logs the event to prompt future outreach.[1][6]

Meta designed the system to distinguish between casual contacts and intimate partners through contextual clues. The agent remembers that a person sending an invoice is a hired plumber, while simultaneously tracking which specific flowers a user's spouse prefers. It builds these profiles to make its automated assistance feel more natural.[2]

The illusion of opting out

The most significant consequence of Muse's architecture is its impact on people who actively avoid Meta's ecosystem. A person can refuse to download the application, decline to upload their data, and opt out of AI training. Yet, their privacy is still compromised if their friends or colleagues use the agent.[6]

Gianluca Miscione, an assistant professor at University College Dublin, highlighted this collateral exposure in a recent analysis. When a Muse user grants the agent access to their inbox and chat histories, every sender in those channels is automatically profiled. The system reads and analyzes messages from highly privacy-conscious individuals regardless of their preferences.[6]

This dynamic creates a network effect where surveillance expands through social proximity rather than direct consent. Investigations found that Muse has compiled lists of vulnerable individuals, including undocumented immigrants and transgender teachers, based purely on their interactions with users. Many of these profiled accounts belong to private citizens with no public persona.[6]

The findings highlight a core tension in the emerging market for autonomous digital assistants. For an agent to successfully book appointments or manage a schedule, it requires deep context about the user's life. However, acquiring that context transforms the tool into a comprehensive surveillance apparatus that affects entire communities.[2][5]

Virtual machines and shared lessons

Meta defends the practice by pointing to the application's underlying security architecture. A company spokesperson stated that all gathered information is stored in a dedicated, secure virtual machine for each individual user. The company maintains that these isolated environments prevent one user's agent from directly accessing another user's personal files.[2]

However, the extracted system instructions reveal that these virtual machines are not entirely siloed. The prompts explicitly state that "Muse agents across many VMs teach each other through shared lessons." The network aggregates behavioral insights to determine which types of nudges and suggestions are most persuasive to human users.[1]

Meta claims this cross-agent communication is strictly anonymized to protect user identities. "To enhance the overall product, we de-identify learnings," a Meta spokesperson stated, confirming that the system scrubs names before sharing behavioral data across the network. The company insists this aggregated data is used solely to improve the software.[1]

Scaling the autonomous agent

Despite the privacy implications, consumers are rapidly adopting the technology. Since its release on September 8, Muse has amassed over 4 million active users and surpassed 5 million total downloads, reaching the number one position in the United States App Store. The software is marketed as a productivity tool.[1][5]

Users are willingly trading their social graphs for the convenience of an automated concierge. Muse can handle restaurant reservations, draft emails, and manage complex travel itineraries on behalf of its owner. To perform these tasks effectively, users routinely grant the application sweeping permissions across their devices and accounts.[2][5]

The scale of this data collection is poised to grow as Meta integrates the agent deeper into its platforms. The company has promised to introduce end-to-end encryption for Muse later this year, which would theoretically prevent Meta itself from reading the dossiers. Until then, the system continues to map the social lives of millions.[1]