AI news · Meta · privacy
Muse instructions tell Meta’s agent to profile a user’s circle
Instructions pulled from the agent show hourly profiles of family, partners, friends and colleagues, with tips on each relationship.
Meta’s Muse agent is instructed to build a profile of each person in its user’s life and to update it every hour, Wired reported on Saturday 3 October 2026. The files came from Karan Joshi, an independent security researcher, who got Muse to hand over its own instructions through the ordinary chat window and passed them to Wired.
The profiles cover family, partners, friends, colleagues, collaborators and people the user follows. Their sections run from facts and history to the relationship, shared interests, open threads and ways to strengthen the tie, and birthdays and anniversaries are among the details kept. Muse is told to record only what it has evidence for, and to leave a page blank before making anything up.
Meta told Wired that an agent needs context about its user and the people around them, drawn from public information and what users choose to share. Each user’s data sits in a dedicated virtual machine, and users can clear what it remembers or unplug connected services. Social Media Today notes that Muse passed 5 million downloads last week and that advertising brings in 97% of Meta’s revenue.
For marketers, the instructions show what a consumer agent is told to record about a buyer’s circle. Social Media Today raises ad targeting as the open question, and neither Meta’s statement nor the instructions address it.
Sources
- Wired: Muse Creates Detailed Profiles of All Your Friends and Family. Checked
- Social Media Today: Meta’s Muse AI collects private user data. Checked
More AI news from
- Altman says OpenAI and Anthropic still split on AI rules
- Trump confirms his Super Intelligence Force on Truth Social
- Musk says SpaceXAI will be renamed SpaceXSI
- GPT-6 Astra swapped in a human-made bot at a StarCraft event
- Ex-Anthropic researcher Coxon testifies to New York City Council
- Lina Khan says AI self-policing repeats a failed playbook
- Claude asks users to share voice recordings for training
- Schneider Electric nears a $20 billion deal for PTC
- DeepSeek cuts the US lead on LiveBench to about 3%