Security researchers at NebuSec say one malicious webpage was enough to take over Meta’s Muse AI agent and send a user’s private chats and stored memories to an outside server, without the approval prompt Muse normally shows. They described the attack in two posts dated 20 November.
NebuSec says it reported the findings to the Muse team and that Meta “had patched all exploits described in this post” by the time it published. We found no public statement from Meta on the research.
Muse is the same agent that Walmart is preparing to connect to its store, according to Marketplace Pulse’s Juozas Kaziukėnas.
Update, 21 November 2026: Muse is also available as a Chrome extension. The Chrome Web Store listing for “Muse from Meta”, published by Meta Platforms, says users can “Open Muse in Chrome’s side panel without leaving the page you’re on” and hand it tasks “from filling out forms to comparing prices and following up on refunds.” The listing says users can “Approve actions before they happen” and that “Muse keeps a full activity log, and lets you revoke any connector anytime.” It showed 300,000 users and version 2.1.12, updated 21 November, when we checked. TestingCatalog reported the extension on 13 November and said Meta had not announced it; we have not seen a Meta announcement either. NebuSec’s findings above concern the Muse agent; NebuSec has not said whether they apply to the Chrome extension.
Business Pill · Least Privilege
An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss this research.
The key insight: NebuSec says one malicious webpage let it pull a Meta Muse user’s chats and memories past the approval prompt and take over the user’s virtual machine; it says Meta patched every exploit before publication.
How the Attack Worked
Muse runs each user’s agent in its own virtual machine, a sealed computer in the cloud. Meta “provisions a 2-vCPU, 8 GiB Cloud Hypervisor VM for each user,” NebuSec wrote, and runs the agent in a container inside it. A gatekeeper called Sentinel watches what leaves.
The way in was the agent’s own web browser. Muse uses a built-in Chromium browser for jobs such as saving pages as PDFs. NebuSec says a single request to export a malicious page gave it code execution inside the agent’s container, which it calls a “0.5-click RCE”.
From there, NebuSec says, weaknesses in how Sentinel inspected outbound requests let it “retrieve conversation history and stored memories without triggering Muse’s network approval prompts”. It says it could also insert fake chat history and alter the agent’s memories.

Three Layers Down, One Held
The researchers then attacked the Linux kernel underneath. NebuSec says its automated exploit tool escaped the container and gained root, full control of the user’s virtual machine, in “more than 99.9% of test runs”.
Root access exposed the machine’s persistent storage, its secure credential storage and Sentinel itself, NebuSec says. It does not say any credentials were taken. By our count that is three layers reached: the agent’s container, the Sentinel gate and the virtual machine.
One layer held. NebuSec says the Cloud Hypervisor boundary, which separates one user’s machine from the shared server, was not crossed, and that it did not reach the shared server or other users’ machines.
Why Agent Security Is Now a Business Question
Agents are being given more to do. One business owner says Muse wrote the $75,000 lawsuit he filed against Meta. Under the planned Walmart connection, Muse would build carts and place orders, Kaziukėnas wrote. On OpenRouter, agent use rose 106 times this year while use by people rose 7 times.
By our arithmetic, agent use grew about 15 times faster than human use on OpenRouter. The Cloud Security Alliance’s daily briefing rated the Muse finding high severity and said “Meta patched every exploit before publication”.
The Structural Read
As we read it, the weak point was the agent’s own browser: the tool that makes an agent useful on the web is also the way in.
As we read it, the layer that held, the hypervisor boundary, is what kept this from reaching other users.
NebuSec, 20 Nov 2026
“Retrieve conversation history and stored memories without triggering Muse’s network approval prompts.”
Three Implications
FOR MUSE USERS NebuSec says Meta has patched every exploit it described.
FOR COMPANIES PLUGGING AGENTS IN Each connection gives the agent more to protect.
FOR AGENT BUILDERS The browser, the outbound gate and the kernel were each a layer NebuSec says it crossed.
What Is Not Established
NebuSec gives no dates for when it reported the flaws or when Meta fixed them, and does not say whether any real user’s data was reached before the fixes. Meta has not commented publicly. The Hacker News also reported the research.
The Bottom Line
NebuSec says one webpage let it pull a Muse user’s chats and memories past the approval prompt and then take over the user’s virtual machine, and that Meta has fixed it. The open question is how agents that browse the web on a user’s behalf are kept from being turned against that user.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The attack details are NebuSec’s own account, from its two posts of 20 November, which we read in full; we have not reproduced the work. The Cloud Security Alliance briefing summarises the same posts. We could read only the headline of The Hacker News report. The count of three layers and the 15 times figure are our arithmetic. We found no public statement from Meta as of 21 November.
Sources: NebuSec, ‘Hacking Muse: Stealing Private Chats Through Browser Use’, 20 Nov 2026 · NebuSec, ‘Inside Meta Muse: Browser, Sentinel and the Shared-Kernel Boundary’, 20 Nov 2026 · Cloud Security Alliance, CISO Daily Briefing, 21 Nov 2026 · The Hacker News, ‘A Single Malicious Webpage Let Researchers Hijack Meta Muse and Access Private Chats’, 21 Nov 2026 · FourWeekMBA on Muse and Walmart, 21 Nov 2026 · FourWeekMBA on OpenRouter agent use, 17 Nov 2026 · FourWeekMBA on a Muse-drafted lawsuit, 19 Nov 2026









