- WIRED and independent researcher Karan Joshi report that Muse’s instructions tell it to build a page for every person in a user’s life, including family, friends, partners and coworkers.
- Experts warn that Muse can guess private details about people who never use Meta, especially when users link their email, calendar and other accounts.
- Meta says Muse works from public information and what users choose to share, and points to safeguards such as isolated virtual machines, permission controls and memory deletion.
Muse, Meta’s new AI assistant, has drawn more than 5 million downloads in three weeks. Now a report raises questions about how it handles personal relationships.
WIRED and independent AI safety and security researcher Karan Joshi say Muse keeps a profile page on every person in a user’s life. Meta says Muse relies on public information and what users choose to share.
How Muse Reportedly Tracks Relationships
Joshi told WIRED he got Muse to reveal its system prompts, the rules that govern how an AI behaves. Other researchers have also shared alleged prompts online. However, we have yet to independently verify the alleged instructions.
According to the prompts, Muse is told to create “a page for every person in the user’s life.” Those people include family, friends, partners, colleagues, collaborators and people the user follows online. FutureSociety researchers say a “Relationships” loop runs every hour in the background on the user’s computer and updates each page.
Muse reportedly builds the pages from messages, photos, and memories of past chats. The page will contain information about the nature of the relationship, their place of residence, their preferences and important dates like birthdays.
The page also includes the relationship’s background, how to contact the individual, and what the user owes or has promised them. Muse can even plan how to strengthen the bond.
Meta’s Response to the Allegation
Meta spokesperson Daniel Roberts told WIRED that an agent needs context about you and the people you deal with to be useful. He said Muse gathers that context from public information and what users choose to share, within the permissions users grant.
Meta’s September 8 article introducing Muse makes it clear that the assistant only uses public information and details that users choose to type in.
Roberts also cited an example during his response to the latest claims. He said Muse can recall that the person who emailed you an invoice is the plumber you hired before.
The Associated Risks
However, where Meta sees a helpful feature, privacy experts see a problem. Carissa Véliz, a professor at the University of Oxford’s Institute for Ethics in AI, notes that AI systems do more than record what users say. They also guess, and some guesses will be wrong. They can mix a user’s data with information from other sources. Those errors can paint a false picture of a person.
Miranda Bogen who directs the AI Governance Lab at the Center for Democracy & Technology also cited some privacy concerns. She says Muse can keep nudging users to link more of their lives, including email, calendars and bank accounts. With that access, the agent can assemble a detailed picture of someone’s routine and relationships.
That puts people outside the app at risk too. Friends, family and coworkers can end up in an AI-made profile without using Muse and without knowing it exists.
Meta’s Safeguards and the Open Questions that Remain
Meta says Muse has several built-in protections. In its September 8 post, Meta AI Research described a dedicated virtual machine for each user. User details and service credentials stay inside that isolated space. A separate permission system reviews each action and asks for approval when needed.
Meta also admits that AI agents make mistakes and can run into malicious instructions in the data they read. It says it uses several layers of defense to limit what an agent can do if someone tries to trick it.
Users can erase stored memories and disconnect outside services. Meta also says users can review the agent’s activity in an audit log.
These tools tackle real safety problems, but a lot of important questions are still up in the air. People want to know exactly what Muse keeps on them, how it makes decisions about others, and how easy it is to view or delete that data.
Meta has also faced legal scrutiny over its privacy representations, including a lawsuit filed by Texas over alleged misleading privacy claims involving Meta and WhatsApp, as reported in Texas Sues Meta and WhatsApp over alleged misleading privacy claims.
There’s an even bigger question underneath all this. Can users actually control what they share and understand the risks? Plus, what about the people mentioned in those notes — do they get any real privacy? Meta needs to clearly answer these questions to put users’ minds at ease.
Share this article
About the Author
Rebecca James is an IT consultant with forward thinking approach toward developing IT infrastructures of SMEs. She writes to engage with individuals and raise awareness of digital security, privacy, and better IT infrastructure.
More from Rebecca JamesRelated Posts
ICANN Switches DNS Root to KSK-2024 in Major Internet Security Update
The root zone KSK rollover took place on October 11, the second time since the first one nearly a de...
Florida Sues TP-Link Over Router Security Risks and Alleged China Links
Florida’s Attorney General James Uthmeier sued TP-Link, saying the company hid risks tied to i...
Microsoft Warns Hackers are Using AI to Speed Up Cyberattacks
Microsoft declared that threat actors are presently reaping AI advantages at a faster pace with rega...
Signal Expands Encrypted Backups and Private iPhone Transfers in Version 8.30
Signal version 8.30 adds on-device backups to iOS and Signal Desktop. The update separates media fil...
Hacker Claims 4.5 Million Records Linked to Italian Hotel Software Firm WuBook Leaked
An underground forum user says they leaked 4.5 million names and IDs tied to WuBook, an Italian hote...
Armenian Ryuk Ransomware Hacker Sentenced to 2 Years in US Prison
Armenian national Karen Vardanyan received 24 months in prison and three years of supervised release...
