Meta's Muse 'dreams' about your chats to guide future interactions
What's the story
A couple of developers have discovered a major vulnerability in Meta's Muse artificial intelligence (AI) platform. Peter James and Jonny L. Saunders independently coaxed Muse into sharing its entire root filesystem, including Ubuntu system files, app templates, and internal documentation. The exploit was achieved with minimal prompting, suggesting that Muse has "almost no prompt injection resistance."
Information exposure
Leak reveals Muse's 'dream' review process
The leak of Muse's filesystem reveals a lot about its internal workings.
It keeps its memory in plain Markdown files and conducts a nightly "dream" review of recent conversations, which is then used to guide future interactions.
The developers also found that many of Muse's features were hard-coded, including one for canceling subscriptions and "the machinery that manages runaway agent spawning."
Feature insights
References to unannounced hardware integration
The leaked files also contained references to hardware integration called Meta Home Link, which seems to give Muse access to devices on a home network.
However, Meta has not officially announced any feature by that name and it is uncertain if it will be implemented.
Company response
Meta defends its security measures
In response to the exploit, Meta has defended its security measures. The company clarified that Muse operates in persistent Linux virtual machines for each user.
A Meta spokesperson, Daniel Roberts, said "Just like with the laptop in front of you, of course you can see the files." He added that exporting virtual machine data doesn't give any privileged access to Meta infrastructure or other people's data.