From The Verge · Terrence O'Brien · · 1m 30s
Meta's Muse AI leaks its own secrets
Developers easily coaxed Meta's new AI into sharing its entire filesystem, revealing how it works and raising questions about its security.
In brief
Developers easily coaxed Meta's new AI into sharing its entire filesystem, revealing how it works and raising questions about its security. Developers extracted Meta's Muse AI filesystem by simple prompting, exposing internal workings. Meta calls it intended behavior but says it poses no security risk. Originally…
Developers extract Muse's secrets

Peter James and Jonny L. Saunders independently coaxed Meta's Muse AI into sharing its entire filesystem, including Ubuntu system files and internal documentation.
Saunders: almost no resistance
Saunders posted that replicating the results was extremely easy and that Muse had almost no prompt injection resistance.
“Almost No prompt injection resistance.”
Meta denies security breach
Meta spokesperson Daniel Roberts said exporting virtual machine data does not give privileged access to Meta infrastructure or other users' data.
“Just like with the laptop in front of you, of course you can see the files.”
Meta calls it intended behavior

Meta Superintelligence Labs' Nat Friedman tweeted this was intended behavior. David Singleton said users should think of Muse as a free computer in the cloud.
“free computer in the cloud”
Files reveal internal workings
The leaked files show Muse stores memory in plain Markdown, performs nightly dream reviews of conversations, and has hard-coded capabilities including subscription cancellation.
Second vulnerability in one week
This is the second Muse vulnerability disclosed this week. Security researcher Patrick Wardle found an exploit that could let attackers hijack the AI agent and access user accounts.
Meta promises ongoing updates

Meta says it is continuing to make updates to the product, so users may see changes in how much information is available about their virtual machine.
What matters most
Developers extracted Meta's Muse AI filesystem by simple prompting, exposing internal workings. Meta calls it intended behavior but says it poses no security risk.




