Muse News

by @tabber

NOV 22, 2026

One webpage let researchers escape Muse's VM and steal chats

Security firm NebuSec says one malicious webpage was enough to take over a single user's Muse agent, pull its chats and memories past the approval gate, and reach root on its cloud VM. Meta patched every flaw before publication, and the hypervisor boundary held.

Reported bynebusec.ai (opens the original story)

In this brief: 3 sections 2 min read
    • Muse's bundled Chromium, used for screenshots, thumbnails and PDF export, lagged behind current Chrome releases and ran a PDF-export process with its IPC sandbox disabled.
    • A public V8 parser bug gave the researchers code execution inside the runtime container.
    • Sentinel's approval rules let read-only requests leave without a prompt, so chats and stored memories were exfiltrated silently.
    • Kernel exploits then escaped the systemd-nspawn container to root on the per-user VM in over 99.9% of test runs.
    • The agent container, Sentinel's egress gate and the per-user VM were all compromised in the lab demo.
    • Root access exposed persistent storage, the credential store and Sentinel itself.
    • The Cloud Hypervisor boundary was not crossed: no shared host or other user's machine was reached.
    • The per-user VM design contained the damage to a single user, which NebuSec calls validation of microVM isolation.
    • NebuSec says it reported every finding to Meta's Muse team and that all exploits were patched before the November 20 publication.
    • The work was a lab demonstration, not an attack on real users, and no dates were given for the report or the fixes.
    • Meta has not posted a public confirmation of the research.
    • The Cloud Security Alliance rated the finding high severity and called it a case study for anyone building autonomous agents that browse the web.
Read full analysis on nebusec.ai ↗
Useful?