Meta's Muse AI Builds Detailed Profiles of Users' Friends
Security researchers have extracted the system prompts of Meta's viral Muse assistant, revealing that the AI compiles hourly, highly detailed dossiers on its users' personal relationships.
Independent AI safety researcher Karan Joshi extracted the system prompts and operating instructions of Meta's new personal assistant, Muse, by asking the chatbot to copy and share its own software files. The leaked documents reveal that Muse, which has already reached millions of downloads, is programmed to run an hourly process that compiles 'a page for every person in the user's life.' This includes structured profiles on family members, partners, friends, colleagues, and social media connections.
According to the instructions, Muse uses structured text files to build these dossiers over time. The profiles can feature sections such as Facts, History, The relationship, In common, Open threads, and Strengthening. The AI is instructed to record specific details like birthdays, anniversaries, resolved arguments, and shared financial goals. It then uses this information to suggest relationship-strengthening actions, such as recommending a coffee shop or prompting the user to make a phone call. Meta spokesperson Daniel Roberts defended the feature, stating that the assistant requires context about a user's social circle to perform helpful tasks.
To secure this sensitive data, Meta designed Muse so that each user has a dedicated virtual machine. This isolated environment prevents other agents from accessing the data, and users can wipe memories or disconnect external integrations at any time. However, privacy experts remain concerned. Carissa Veliz, an associate professor at Oxford's Institute for Ethics in AI, warned that users are giving AI systems far more information than they receive. Miranda Bogen of the Center for Democracy and Technology noted that Muse places a heavier emphasis on personal contacts than its competitors, encouraging users to proactively share extensive personal data.
For AI developers and product designers, the Muse leak highlights a major shift toward hyper-personalized, agentic systems that rely on continuous background profiling. While dedicated virtual machines offer a technical safeguard for data isolation, building systems that actively solicit and map out real-world social graphs introduces massive privacy liabilities. Practitioners must balance the competitive advantage of deep personalization against the growing regulatory and ethical risks of managing highly sensitive, interconnected user data.
This is our own summary of reporting by WIRED AI
