One Webpage Hijacked Meta Muse and Reached Its Private Chats

Human Verified

Share to: , and

A laptop showing a glowing AI assistant while a red strand of light reaches from a cracked malicious webpage to an open filing cabinet drawer

A single malicious webpage could hijack Meta Muse, pull a user’s private chats and stored memories to an outside server, and then take root in the cloud machine running the assistant. Security researchers published the attack chain on 20 November, after reporting the findings to Meta.

How the Attack Worked

The attack starts with Muse’s built-in browser, which the assistant uses for jobs like turning webpages into PDFs and other documents. An attacker-controlled page exploited a flaw in that browser to run code inside the agent’s runtime container, the first quiet foothold in a longer chain.

From there, NebuSec found weaknesses in Sentinel, the service meant to police what Muse sends out. The researchers showed conversation history and stored memories could leave quietly for an external server with no approval prompt, and could also be altered on the way.

From Browser to Root

The chain did not stop at the container. Using its own automated exploit generator, NebuSec built Linux kernel exploits that broke out of the container and won root privileges in the user’s cloud machine, succeeding in more than 99.9 percent of test runs.

Root access exposed the parts Muse keeps walled off from the agent: secure credential storage, all saved files, and Sentinel itself. The compromise stayed inside one user’s machine, with no escape to the shared host beneath it or to other users’ machines.

Our Read

The striking part is that every single control worked exactly as designed and the attack still ran end to end. Sandboxing, privilege separation, credential isolation and network monitoring each did their job, yet a chained attack crossed all four boundaries in sequence.

That is the lesson NebuSec draws for every team building agents with real permissions. Muse has already faced trust questions over invented bank balances in Signal chats, and this demonstration shows the risk is not only what an agent says, but what an attacker can make it do.

Meta has not publicly confirmed the fixes or commented on the research, and the write-up does not say whether any user data was reached before the patches went in. NebuSec says all the exploits were patched before publication, which limits the exposure but leaves the silence.

Why It Matters

Muse is Meta’s personal AI agent, launched on 8 September, and it works through the accounts a user connects, with permission to make purchases on their behalf. An agent that can spend money and read private conversations is a high-value target for exactly this kind of attack.

The finding also travels beyond Meta. Every frontier lab is shipping agents that browse, run code and hold credentials, and the same chain — a webpage flaw, a network-policy gap, a kernel escape — could apply to any of them. The industry’s sandboxing assumptions now face a public test.

Background

This is the second serious Muse flaw in three months. An earlier September flaw let malware already on a user’s Mac hijack the assistant on the device, but that one never reached the cloud system. Meta keeps expanding Muse, including a $300 companion device with free Verizon 5G.

Source: NebuSec research and The Hacker News

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.