Single Malicious Web Page Allows Researchers to Hijack Meta Muse and Access Private Chats

Swati KhandelwalNovember 21, 2026Vulnerability / Artificial Intelligence

A malicious web page could extract a Meta Muse the user’s private conversations and the assistant’s stored memories without triggering any approval prompts, the researchers showed.

The same attack could be extended to reach credentials stored by Muse for the user’s logged-in accounts.

The work was a lab demonstration, not an attack on real users, and it remained inside each user’s cloud machine. NébuSecthe security company that released it yesterday, claims that the Muse team has since fixed all of the weaknesses described.

Muse is Meta’s personal AI agent, launched September 8, that works through the online accounts a user logs into and can make purchases on their behalf. Each person’s agent runs on their own cloud computer.

There, a monitoring layer called Sentinel is supposed to be the single door for everything the agent sends, and passwords and tokens for connected accounts are supposed to stay out of reach of the agent.

The Muse team described the goal of the launch: “So we designed the system to assume that the agent could be attacked and limit the potential damage.”

In the demonstration, a malicious page opened by the agent could extract a user’s chat history and stored memories on an external server and modify them, without any prompts being displayed to the user, according to the company’s research report.

From there, the researchers chained a flaw in Muse’s built-in browser, which the assistant uses for tasks like creating PDFs, to a flaw in the Linux kernel, effectively taking full control of the user’s cloud machine. They presented this channel in a technical outage.

This total control reached the credential store, saved files, and Sentinel itself, the elements that Muse’s security design keeps away from the agent.

The compromise went no further than this one machine. The researchers did not penetrate the shared server below or other users’ machines.

Muse was designed to assume that its own agent could turn against the user and to prevent any break-in. Each of these controls worked as expected, the company said, but the attack continued until the end.

The company says it reported the findings to the Muse team and that the exploits were all fixed before release. Meta has not publicly confirmed the fixes or commented on the work, and the text does not specify whether any user data was accessed before the fixes arrived.

The company’s Muse research penetrated the cloud system behind the assistant, unlike an earlier Muse flaw discovered in September, which only allowed malware already on a user’s Mac to hijack the assistant on the device.

Gn bussni

Logo