One bad web page gave researchers control of a Meta Muse machine
NebuSec tells that a bad web page could send private chats from Muse with no approval step, and that the weak points are repaired.
The security firm NebuSec wrote a report about a test against Muse, a personal AI agent that Meta launched on September 8. A bad web page could send the private chats and memories of a user to an external server, with no approval step. NebuSec tells that the team of Muse repaired all the weak points before the report. Meta did not tell if the repairs are correct.
How it works
Muse works through the accounts that a user connects to it. Each agent operates on its cloud machine, and a monitor layer, Sentinel, must check all data that goes out.
The researchers connected a weak point in the browser of Muse to a weak point in the Linux kernel, the primary part of that system. This gave them full control of the machine, the passwords of the connected accounts and Sentinel.
The proof
This was a test, not an attack on users. The researchers did not go to the shared server or to the machines of other users. NebuSec tells that each safety control operated as it must, but the attack went through.
What is not known
Meta did not tell its view of the work. The report does not tell if a person got data of users before the repairs.
Insight: External tests keep finding holes in Muse
Meta told that Muse is more safe than OpenClaw, and Mark Zuckerberg told that it is made for privacy. Since then, external researchers found a zero-day weak point, share links that showed more than 6,800 chats, and this attack. This reads as external groups, not Meta, that set the safety record of Muse. Before staff connect work accounts, check the results of tests by a different group.
Sources
Posted