Malicious Webpage Hijacked Meta Muse, Stole Chats

Security researchers have shown that a single malicious webpage could hijack Meta Muse, the company's personal AI assistant, and pull a user's private conversations and stored memories out to an outside server without showing any approval prompt. The research, published by security firm NebuSec on 20 November and reported on 21 November, also found the same attack could be extended to reach the login credentials Muse stores for the user's connected accounts.
Muse launched on 8 September 2026 as Meta's personal AI agent. It works through the online accounts a user connects to it and can make purchases on their behalf. Each person's agent runs on its own cloud computer, and Meta built the system around the idea that the agent itself might be turned against its user. A monitoring layer called Sentinel is meant to act as the single gate for anything the agent sends out, while the passwords and tokens for connected accounts are supposed to stay beyond the agent's reach.
How the attack chain worked
The attack began with a flaw in the Chromium browser Muse uses for everyday tasks such as creating PDFs and documents. By getting the agent to open an attacker-controlled webpage, the researchers achieved code execution inside the container where the agent runs. At that point Muse's controls still limited outside network access, but NebuSec's analysis of Sentinel found weaknesses in how outbound requests were classified and inspected. The researchers showed that conversation history and stored memories could be sent to an external server without triggering an approval prompt, and that the stored memories and conversation history could also be altered.
From there the researchers chained the browser flaw to flaws in the Linux kernel, using their own automated exploit-generation framework to build kernel exploits that escaped the container and gave them root privileges in the underlying user machine. That step exposed the resources Meta had deliberately kept outside the agent's runtime environment, including persistent storage, the secure credential store and Sentinel itself. Once the machine was under their control, outbound network traffic was no longer subject to Sentinel's enforcement at the container level. NebuSec said its kernel exploits succeeded in more than 99.9 percent of test runs.
Contained to one machine, but the lesson is wider
The compromise went no further than that one machine. The researchers did not break out to the shared server beneath it or into other users' machines, and they stressed that the work was a laboratory demonstration, not an attack on real users. It differs from an earlier Muse flaw found in September, which only let malware already on a user's Mac hijack the assistant on that device. This time the researchers broke into the cloud system behind the assistant.
NebuSec said it reported the findings to Muse's team and that every weakness it describes had been fixed before publication. Meta has not publicly confirmed the fixes or commented on the work, and the write-up does not say whether any user data was accessed before the patches went in. The firm also noted that Muse's controls worked as designed at each step, yet the attack still ran the whole way through, because each stage weakened a different part of the security architecture.
What users should do
For Muse users, the practical message is that the demonstrated flaws are patched, according to the researchers, and there is no evidence that real users were attacked. Even so, the episode is a reminder to check which accounts an AI assistant is connected to, to remove connections that are no longer needed, and to keep the assistant and its apps updated. Anyone who uses an AI agent that can browse the web, run code and act on their accounts should treat the pages it opens as untrusted, because a single hostile page was enough, in this demonstration, to reach everything the agent was supposed to keep out of reach.