Meta Muse: Malicious Webpage Exposed Chats in a Test
On November 20, 2026, security firm NebuSec published research showing that a malicious webpage, exported by Meta's Muse agent, let the researchers pull chats and memories without going through Muse's approvals. It was a lab test and has been patched. It matters to anyone using AI agents with a browser and connected accounts.
What we know
- What Muse is: Meta's cloud-hosted personal AI agent, running in a virtual machine dedicated to each user. It can browse with its own browser, run code and connect to services such as email or calendar, according to NebuSec and Meta's own blog.
- The entry point: according to NebuSec, the chain began with a flaw in the Chromium browser Muse uses to generate PDFs and documents. By asking the agent to export a page they controlled, the researchers ran code inside the agent's container.
- What they could take: NebuSec says conversation history and stored memories were sent to an external server without triggering Muse's network approval prompt. They could also modify those memories and conversations.
- How far they got: using a Linux kernel flaw, they escaped the container and gained administrator (root) rights in the user's virtual machine, reaching the secure credential store and Sentinel, the Meta service that polices outbound traffic.
- The limit: the compromise stayed inside that user's virtual machine. NebuSec says it did not demonstrate a jump to the host server or to other users, and there is no indication it was used against real users.
- The fix: NebuSec reported its findings to the Muse team and says every exploit it describes had been patched by publication time. According to The Hacker News, Meta has not commented publicly on the findings.
- Meta's design: in its September blog post, Meta describes Muse as having layered isolation, a service that authorizes every action and outbound connection, credentials kept out of the agent's reach and a bug bounty program. NebuSec chained three flaws to get across those layers.
What changes and what doesn't
If you use Muse today, there is nothing to update: according to NebuSec, the flaws are already fixed in the service.
What changes is the assumption that approvals are enough. In this test, the user only asked for something ordinary (saving a page) and data left without a prompt. What we don't know is whether other browsing agents have similar flaws: the research only looked at Muse.
How to know if this affects you
If you use an AI agent that browses for you and has access to your accounts (Muse or another), or you plan to put one to work in your business, check this:
- Be careful what you hand it: don't ask it to open, save or export pages from links you don't know, especially ones that arrive by email or message. In the test, that was the only step needed.
- Connect only what's essential: every connected account (email, calendar, payments, social) is one more thing to lose if the agent is compromised. When testing an agent, use separate accounts with no customer data.
- Don't rely on approvals alone: set it to ask before every sensitive action, but assume a flaw can bypass the prompt. Anything you can't afford to lose shouldn't be within the agent's reach.
- Review its memories now and then: the test showed they could be altered. If your agent keeps memories, look through them regularly and delete anything you don't recognize or that it shouldn't know.
- If you build your own agent (for customer service or internal tasks): keep its browser up to date, store credentials outside where the agent runs and give them the narrowest scope, which is what NebuSec recommends to agent builders.
Related: Signal Blocks Screen Capture by iOS 27 and Android AI
Sources
- NebuSec, «Hacking Muse: Stealing Private Chats Through Browser Use», November 20, 2026
- NebuSec, «Inside Meta Muse: Browser, Sentinel and the Shared-Kernel Boundary», November 20, 2026
- Meta, «How We Built Safety Into Muse», September 8, 2026
- The Hacker News, November 21, 2026
Updates: if Meta publishes a response or details of the fix, it will be added here with a link.