News

A Single Malicious Webpage Let Researchers Steal Meta Muse Chats and Account Credentials

NebuSec showed a malicious page Muse exported as a PDF could steal chats and memories, then reach credentials in the user's cloud VM. The firm says Meta patched it.

Meta built Muse so a compromised agent would stay boxed in. Researchers asked it to export a hostile webpage as a PDF, then read the user's private chats from the same machine.

NebuSec published the work on November 20. It was a lab demonstration, not an attack on real users, and it stayed inside one person's cloud VM. The firm says it reported the chain to Muse's team and that every exploit in the write-up had been patched by publication. Meta has not posted a public confirmation.

Muse launched on September 8 as a personal agent that can use the accounts a user connects and, in Meta's pitch, make purchases. Each user gets a dedicated cloud computer. A monitoring layer called Sentinel is supposed to gate anything the agent sends out. Passwords and tokens for those connected accounts are supposed to sit beyond the agent's reach.

A PDF export, then a silent GET

The opening move used Chromium bundled beside the agent for jobs such as generating PDFs. In a technical breakdown the same day, NebuSec said that PDF-export process had its IPC sandbox off. When a user asked Muse to save an attacker-controlled page, a renderer bug was enough for code execution inside the runtime container, with the agent's privileges.

Container root still could not talk to the internet on its own. Sentinel sits outside that container, rewrites outbound connections, and decides which requests need a prompt. Under Muse's default setting, only GET, HEAD, and OPTIONS requests with no body counted as reads, and those could leave without asking. The researchers used that path to send conversation history and stored memories to their server with no approval dialog. They also altered those memories and injected chat entries.

That already crossed the line Muse advertised: the agent was supposed to be under attack, and the damage was supposed to stay limited.

Then the box around the box

The next step left the container. NebuSec's exploit-generation framework produced Linux kernel exploits against the custom Ubuntu kernel in the per-user VM. In the firm's tests those exploits succeeded more than 99.9 percent of the time. VM root exposed persistent storage that was not mounted into the container, the credential store, and Sentinel itself. Processes running directly in the VM were not forced through Sentinel's cgroup hook, so outbound traffic no longer needed a prompt.

The break stopped there. The researchers did not get out of Cloud Hypervisor to the shared host or into another tenant's machine.

Cyberpresso already covered a different Muse hole, in dictation. This one starts with a page the agent is asked to save. The architecture did what it was designed to do at each layer, and the chain still ran the length of one user's VM.

Some offers on this page may be paid placements or contain affiliate links.

Cyberpresso: daily cyber & AI brief

Free daily newsletter, read in 5 minutes.

Subscribe free