The Pulse
Meta Muse’s Hidden Instructions Map Users’ Relationships
Meta’s Muse includes instructions to build a page for each person in a user’s life, drawing on relationship details and other information available to the assistant. The discovery comes as millions use the agent and connect it to services i
AI.info Team · Written with AI ·
“They’re trying to know you like a friend, which is honestly pretty creepy.”
Karan Joshi, independent AI safety and security researcher, speaking to WIRED
Muse’s instructions call for a page for each person
Meta’s Muse is designed to build detailed records about people in a user’s life, according to internal instructions obtained by a researcher and reviewed by WIRED. The files direct the assistant to create “a page for every person in the user’s life,” including family members, partners, friends, colleagues, collaborators, and people the user follows.
The pages draw on Muse’s memory, which WIRED describes as structured text files. The instructions envision records that can include facts, relationship history, shared interests, ongoing conversations, and ways to strengthen a relationship. The assistant could also retain dates such as birthdays and anniversaries, or remember a friend’s coffee preference when suggesting where to meet.
The finding describes what Muse is instructed to do; it does not establish that the system has already filled out a complete profile for every person connected to every user. The documentation says a page may start sparse and expand over time, and directs the model to rely on available evidence rather than invent details.
Researchers extracted files through Muse’s chat interface
Independent AI safety and security researcher Karan Joshi extracted Muse’s instructions and system prompts by asking the agent through its regular chat interface to copy and share its software files. He shared the findings with WIRED. Meta has said the files were accessible intentionally as part of its approach to transparency, according to the report.
The instructions describe relationship records in unusually specific terms: how close two people are, what their relationship is built on, how they interact, and what it might need. Suggested follow-ups could include a reason to call, something a person said to revisit, or an opportunity to offer support.
That design gives the agent a basis for personalized suggestions, but it also turns ordinary conversations and connected services into potential sources of information about third parties. A person whose details appear in a user’s messages may have no direct relationship with Muse, yet those details can inform its view of the user’s relationships.
Meta says Muse uses public information and user-shared data
Meta spokesperson Daniel Roberts told WIRED that the agent needs context about users and the people they interact with to be useful. “Muse gathers that based on public information and from what you’ve chosen to share,” Roberts said, citing examples such as recognizing a plumber who previously worked in a user’s home or recalling flowers a spouse liked.
Meta’s system design gives each user a dedicated virtual machine to store data and context; WIRED reports that the environment is inaccessible to other agents. Users can wipe memories or disconnect outside services, and the company says Muse asks for confirmation before actions such as sending email or making purchases. An audit log shows the agent’s activity and planned actions.
Personalization also widens what the agent can learn
Carissa Véliz, an associate professor at Oxford’s Institute for Ethics in AI, told WIRED that AI systems can learn from what people state directly, what they infer, and information they assemble from other sources. She warned that inferences can be wrong as well as concerning.
Miranda Bogen, director of the Center for Democracy and Technology’s AI Governance Lab, said Muse appears to place more emphasis on relationships and personal contacts than competing assistants. She also pointed to the breadth of information these agents invite people to connect, including email, calendars, and financial institutions.
Muse’s proposed relationship pages make that tradeoff tangible: a memory file could combine where someone lives, what they do, a recurring shared concern, and a date worth remembering. The instructions tell the system to leave gaps rather than fill them with guesses, but the accumulated record can still reach well beyond what a user explicitly enters as a profile.