Instructions extracted from Meta Muse describe an hourly process for building or updating memory pages about people in a user’s life, including people who have never used the agent, WIRED reported. The finding documents a possible design, not proof that every account creates files on every contact or that the process runs hourly in practice.
That distinction shapes the privacy concern: details about a non-user could be kept in another person’s Muse environment. The available checks have not established what populated files contain across accounts, how long they persist or whether people described in them can see or delete them.
What could Meta Muse’s reported files contain?
WIRED reported Saturday that independent AI safety and security researcher Karan Joshi obtained Muse instructions and system prompts through its regular chat interface, then shared them with the outlet. The instructions describe pages for family members, partners, friends, colleagues, collaborators and people a user follows.
Those pages could organize details such as where someone lives and works, recurring matters, birthdays and anniversaries, shared history, relationship context, and possible reasons to reconnect or offer support. The instructions describe an hourly schedule, but that does not confirm how often the process runs in deployed accounts or whether every listed detail is stored.
WIRED’s report does not establish that each user has a populated page for every person they know. Meta announced Muse as a personal AI agent on Sept. 8; WIRED said the pages are intended to give the system context for personalized suggestions involving people in a user’s plans.
How was the reported behavior checked?
Joshi’s method showed which instructions Muse exposed through that chat route. It did not establish the behavior of background processes or confirm what files contain across users’ accounts.
On Tuesday, Amanda Caswell, AI editor at Tom’s Guide, asked her own Muse what it knew. Her account of the check describes some details about her household, but thinner, unnamed information about people outside it. Caswell said she had connected relatively little data.
That single-user check offers a limited view of one Muse account; it is not an audit of stored files or a test across multiple accounts. Its sparse results do not establish what other users’ Muse instances retain.
What does Meta say about the information and controls?
Meta spokesperson Daniel Roberts told WIRED that Muse uses public information and details users choose to share to build context about people they interact with. That is Meta’s explanation of the system’s purpose and inputs, not independent confirmation of what is stored or how the reported process works in practice.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
Meta says each Muse user has a dedicated virtual machine for their data and context. It also says users can inspect, edit and download files in that environment, including memories about themselves, and can wipe memories or disconnect external services. The available reporting does not establish whether a person mentioned in another user’s files can inspect or remove information about themselves directly.
What does this mean for people who have never used Muse?
If the described pages are created, information about someone could sit in another user’s Muse environment even if that person never signed up. The privacy issue is therefore not limited to what a system gets from a person’s own account: Meta’s spokesperson said Muse also draws on public information and details shared by users.
Carissa Véliz, an associate professor at Oxford’s Institute for Ethics in AI, told WIRED that concerns also include information AI systems may infer from other data, whether those inferences are right or wrong. Her warning is a broader concern, not evidence that a particular Muse page contains an inaccurate or sensitive inference.
The open questions are whether the described hourly jobs run in deployed accounts, what actual pages contain and how long they persist. The reporting also does not establish whether non-users can find out if they are represented or request deletion themselves, or whether independent researchers have examined populated pages across multiple accounts.