A malicious webpage might extract a Meta Muse person’s non-public conversations and saved reminiscences from the assistant with out triggering any approval immediate, researchers have proven.
The identical assault may very well be prolonged to succeed in the credentials Muse shops for the person’s linked accounts.
The work was a laboratory demonstration, not an assault on actual customers, and it stayed inside the person person’s cloud machine. NebuSec, the safety agency that revealed it yesterday, says Muse’s workforce has since fastened each weak spot it describes.
Muse is Meta’s private AI agent, launched on September 8, that works by way of the web accounts a person connects to it and may make purchases on their behalf. Every individual’s agent runs by itself cloud laptop.
There, a monitoring layer referred to as Sentinel is supposed to be the one gate for something the agent sends out, and the passwords and tokens for linked accounts are supposed to remain past the agent’s attain.
Muse’s workforce described the intention at launch: “So we designed the system to imagine the agent could also be underneath assault and restrict the potential injury.”
Within the demonstration, a malicious web page the agent opened might pull a person’s dialog historical past and saved reminiscences to an exterior server and alter them, with no immediate proven to the person, in line with the agency’s analysis write-up.
From there, the researchers chained a flaw in Muse’s built-in browser, which the assistant makes use of for duties equivalent to creating PDFs, to a Linux kernel flaw, thereby taking full management of the person’s cloud machine. They set out that chain in a technical breakdown.
That full management reached the credential retailer, the saved recordsdata, and Sentinel itself, the components Muse’s safety design retains walled off from the agent.
The compromise went no additional than that one machine. The researchers didn’t get away to the shared server beneath it or into different customers’ machines.
Muse was constructed to imagine its personal agent may very well be turned in opposition to the person and to field in any break-in. Every of these controls labored as designed, the agency stated, but the assault nonetheless ran the entire means by way of.
The agency says it reported the findings to Muse’s workforce and that the exploits have been all fastened earlier than publication. Meta has not publicly confirmed the fixes or commented on the work, and the write-up doesn’t say whether or not any person information was accessed earlier than the patches went in.
The agency’s Muse analysis broke into the cloud system behind the assistant, not like an earlier Muse flaw present in September, which solely let malware already on a person’s Mac hijack the assistant on the machine.








