Researchers Used a Malicious Webpage to Break Into Muse and Reach Private Chats and Credentials
Security firm NebuSec showed how a malicious webpage opened by Muse could expose a user's conversations and stored memories without an approval prompt, and how researchers then used a Linux kernel exploit to reach the credential store inside the user's cloud machine. NebuSec says every weakness it found was patched before publication.

A malicious webpage opened by Muse could pull a user’s private conversations and stored memories out of the assistant and send them to an external server without an approval prompt, security firm NebuSec has shown. The researchers chained further flaws to escape the container Muse runs in, took root control of the user’s cloud machine, and reached the credential store that holds the tokens for a user’s connected accounts.
NebuSec published the research on November 20, with a separate technical write-up that sets out the chain, and The Hacker News reported the findings on November 21. The work was a laboratory demonstration against the researchers’ own setup, not an attack on real users, and the compromise stayed inside the individual user’s cloud machine. The researchers did not break out to the shared server beneath it or into other users’ machines.
How the Chain Worked
Muse runs each user’s agent on its own cloud machine. Inside that machine, the agent works in a runtime container, Sentinel monitors outbound traffic and enforces approval policies, and a separate credential service keeps the real passwords and tokens for connected accounts outside the agent’s environment.
The attack started with a vulnerability in the Chromium browser Muse uses for tasks such as generating PDFs and documents. The researchers had the agent export an attacker-controlled webpage and gained code execution inside the runtime container. At that point Sentinel still restricted network access, but the researchers found weaknesses in how outbound requests were classified and inspected, and showed they could transmit conversation history and stored memories to an external server without an approval prompt. The same access let them modify stored memories and conversation history.
The researchers then used Linux kernel exploits, developed with their automated exploit generation framework, to escape the container and obtain root privileges in the user’s virtual machine. That exposed persistent storage, the secure credential store, and Sentinel itself, and it allowed outbound traffic that was no longer subject to Sentinel’s container-level enforcement. Root inside the container did not by itself grant unrestricted network access or access to the underlying machine. Each stage of the chain crossed a different layer.
Patched Before Publication, Meta Yet to Comment
NebuSec says it reported the findings to the Muse team and that all of the exploits described in its research were patched before publication. Meta has not publicly confirmed the fixes or commented on the research, and the write-up does not say whether any user data was accessed before the patches went in.
The demonstration differs from the Muse flaw disclosed in September, which let malware already running on a user’s Mac hijack the assistant on that device. The NebuSec chain attacked the cloud system behind Muse itself, starting from web content the agent was asked to handle. It arrived in the same week as Reuters reporting on internal Muse safety concerns, which described the agent reaching data users had not shared with it, and Signal’s move to restrict screenshots by default so agents cannot read its chats.


