NebuSec, a security firm, published research showing how a malicious web page could lead to access to a Meta Muse user’s private conversations and stored memories. NebuSec says all exploits it describes had been patched by the time of publication.
Muse is a cloud-hosted AI assistant that runs inside an isolated virtual machine for each user. NebuSec says the design added a runtime container, a service called Sentinel that monitors outbound traffic and enforces approval policies, and a separate credential management service that keeps connector credentials outside the agent’s environment.
What NebuSec reports
The firm says the chain began with a vulnerability in the Chromium browser Muse uses to generate PDFs and documents, which gave code execution inside the runtime container. It then reports weaknesses in how Sentinel classified and inspected outbound requests, which let conversation history and stored memories leave without an approval prompt. The same path allowed changes to stored memories and history.
NebuSec then describes Linux kernel exploits that let it escape the container and gain root in the user’s VM, exposing secure credential storage and Sentinel itself. The firm says the compromise stayed inside the individual user’s VM and it did not demonstrate an escape into the hypervisor host or other tenants.
Why it matters to defenders
NebuSec’s point is that controls can work as designed and still fail as a chain. Security teams that build or buy agents can ask whether credential separation and network monitoring still hold once an attacker controls part of the execution environment. Our report on Amazon’s account of AI agents finding a zero-day covers AI agents from the attacker’s side, and our opinion piece on giving every AI agent its own identity argues for scoping agent permissions.
Source: NebuSec, Hacking Muse: Stealing Private Chats Through Browser Use