Security firm NebuSec says it chained Chromium and Linux kernel bugs to escape the sandbox around Meta's Muse AI agent, starting from a single malicious webpage and ending with root access on the agent's cloud virtual machine. NebuSec published the research on Nov. 20 and says the Muse team was aware of every finding and had patched all of the exploits before publication.
Muse, which Meta launched on Sept. 8 as a consumer personal agent, gives each user a 2-vCPU, 8 GiB Cloud Hypervisor VM. Inside it, the agent runs as root in a systemd-nspawn container, with a separate service called Sentinel acting as the gatekeeper for network traffic. Meta's launch materials said nothing Muse does reaches the internet unless Sentinel approves it.
The entry point was Muse's bundled Chromium browser, which the agent uses for screenshot review, thumbnails and PDF export. In early October that build was still Chromium 152.0.7977.82, a patch release from Sept. 3, leaving more than five public V8 vulnerabilities usable, according to NebuSec. The firm used a V8 parser bug that escaped the V8 sandbox to get code execution.
Key facts
- 01NebuSec published research on Nov. 20 showing a malicious webpage could compromise Meta's Muse agent and reach root on its VM.
- 02Muse's bundled browser ran Chromium 152 in early October, leaving more than five public V8 bugs usable, per NebuSec.
- 03NebuSec exfiltrated a test user's chat history and memories past Muse's Sentinel egress gate with no approval prompt.
- 04NebuSec says its kernel exploits succeeded more than 99.9% of the time; it did not show a hypervisor escape.
- 05NebuSec says Meta's Muse team patched all described exploits before publication; Meta's bounty pays up to $300,000.
Because the PDF-export browser ran with its IPC sandbox disabled, one renderer bug, which NebuSec valued at about $500, was enough to run shell commands inside the agent's container. NebuSec described the result as a "0.5-click" remote code execution.
Getting data out meant beating Sentinel. NebuSec found that bodiless GET, HEAD and OPTIONS requests count as reads, which can leave without a prompt under Muse's default "Ask for some actions" setting. Sentinel screens URLs for emails, phone numbers and verification codes, decodes hex and Base64, flags high-entropy chunks and routes suspicious requests to a model that can veto them. A short Base64-encoded email triggered review, while the same address in plaintext passed.
Using that path, NebuSec says it sent a test user's conversation history and stored memories to its own server through the browser chain with no approval prompt. It could also inject history entries and alter memories.
The firm then went after the VM's kernel, a Hatch-built Ubuntu 7.0.0-39-generic build compiled Sept. 24 with several hardening options, including RANDOMIZE_MEMORY and HARDENED_USERCOPY, disabled. NebuSec says its automated exploit generation framework produced kernel exploits that succeeded more than 99.9% of the time.
From VM root, NebuSec says it could read the user's 100 GiB encrypted volume, Muse's Secure Credential Storage, Sentinel itself and the boot partition. Processes running directly on the VM sat outside the container's network hook and had their own public IPv6 address. NebuSec says it did not demonstrate a hypervisor escape or access to other users, and the testing ran against its own Muse instance.
The Cloud Security Alliance's AI Safety Initiative, which analyzed the findings on Nov. 21, said none of its sources included a Meta statement on the NebuSec research and attributed the patch timeline to NebuSec. Meta opened its bug bounty to everyone on Muse's launch day, paying up to $300,000 for valid reports, including up to $130,000 for a successful prompt injection affecting one user.
The research follows a Nov. 18 report that Muse retrieved Signal message content on Android and macOS during tests with Signal staff and the AI Now Institute, using OS permissions granted at setup. A Meta spokesperson described those cases as isolated incidents.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.




