Innovation & Tech

Meta Muse compromised in lab via one malicious webpage

NebuSec says one malicious webpage let researchers escape Meta Muse’s cloud container, reach VM root and access chats and memories. All reported flaws were patched before publication.

By Grace Kim

4 min read

Meta's Muse AI Agent Compromised via Single Malicious Webpage in Lab Demonstration
Meta's Muse AI Agent Compromised via Single Malicious Webpage in Lab Demonstration377053 / pixabay

What's News

  • NebuSec's container escape succeeded in more than 99.9% of test runs.
  • Muse launched September 8, 2026 in a dedicated 2-vCPU, 8-GiB cloud VM.
  • The chain used Chromium 155.0.8059.80 and Ubuntu kernel 7.0.0-39-generic.
  • NebuSec reported the flaws to Meta and says all exploits were fixed before publication.
  • The lab compromise reached one user's VM, not the hypervisor or other tenants.

By Grace Kim

NebuSec researchers said a single attacker-controlled webpage compromised Meta's Muse personal AI agent in a lab demonstration, with their container escape succeeding in more than 99.9% of test runs. The chain reached private conversation history, stored memories, saved files, the credential store and Sentinel, Meta's outbound-request monitoring layer.

NebuSec described the work as a lab test against one user's cloud machine, not an attack on real users. The researchers said they reported all vulnerabilities to Meta and that all exploits were fixed before publication. Meta has not publicly commented on the findings or confirmed the patches.

How did one webpage reach VM root?

Meta launched Muse on September 8, 2026. Each user's agent runs in a dedicated cloud VM with 2 vCPU, 8 GiB of memory and Cloud Hypervisor, inside a systemd-nspawn runtime container.

The agent operates as root inside the container, but its user namespace maps UID 0 to unprivileged host UID 131072. The hatch-authd credential store, Sentinel and safety classifiers run outside the container in the user VM, according to NebuSec.

The researchers started when Muse opened a malicious page in its bundled Chromium browser, version 155.0.8059.80 at testing. Muse uses the browser for PDF export and thumbnail generation. A V8 sandbox-escape bug, issue 554034656, delivered arbitrary code execution inside the runtime container.

NebuSec's result turned a routine browsing action into the first stage of code execution.

How did Sentinel fail to stop the data transfer?

NebuSec said Sentinel classifies outbound requests by process origin, HTTP method, content entropy and cumulative traffic budgets. Writes require approval unless an allow rule matches. Reads using GET, HEAD or OPTIONS without a body may leave silently under the default “Ask for some actions” setting.

By crafting exfiltration requests that Sentinel treated as reads, the researchers sent conversation history and stored memories to an external server without triggering a user prompt. They also injected into and altered entries in the conversation history and memory store, NebuSec said.

In practical terms, the default policy did not interrupt a read-classified theft.

How did the researchers escape the container?

From the compromised container, the team targeted a shared custom Hatch-built Ubuntu kernel, version 7.0.0-39-generic. It ran with RANDOMIZE_MEMORY, HARDENED_USERCOPY and INIT_ON_ALLOC_DEFAULT_ON disabled, while CONFIG_IO_URING was enabled but blocked at runtime.

Using an automated exploit-generation framework, the researchers developed a container escape that reached root in the user VM in more than 99.9% of test runs, without requiring a target kernel image.

VM root gave them access to the already-mounted LUKS2/Btrfs persistent volume, the secure credential store, Sentinel and the host's public IPv6 interface. The IPv6 route bypassed Sentinel's eBPF-enforced proxy, according to NebuSec.

The compromise stopped at the user VM and never reached the hypervisor or other tenants' VMs, the researchers said.

Which controls held, and where did they fail?

Meta's launch blog describes a “defense in depth” architecture. It isolates the agent in a runtime container and keeps real OAuth tokens and passwords beyond the agent's reach in hatch-authd. The system substitutes surrogate tokens that Sentinel swaps for real credentials at the network boundary.

Sentinel serves as the sole gate for connector actions and network egress. Meta's “tainted egress” system tracks whether a process has read user data, then determines whether to auto-allow an action or prompt the user.

NebuSec's analysis found that the container mapping held and surrogate tokens never reached the agent. Sentinel also initially blocked unauthorized exfiltration. The browser exploit nevertheless established an in-container foothold, read-classification logic permitted the transfer, and the kernel escape crossed the remaining boundary.

Each layer worked within its stated design, yet the chain reached data the architecture sought to isolate.

Were real users affected, and what remains unresolved?

NebuSec said the demonstration did not establish whether anyone accessed real user data before the fixes. The researchers' write-up does not resolve that question, and Meta has issued no statement about the research or the reported patches.

An earlier flaw disclosed in September 2026 allowed malware already present on a user's Mac to hijack the local Muse assistant. The new chain differed because it broke into the cloud backend itself, according to NebuSec.

The risk shifted from a compromised local device to the service's cloud isolation layer.

What does this mean for consumer AI agents?

Muse is among the first shipping consumer agents to hold delegated credentials for email and calendar, plus payment via Stripe Link, while running a full browser in a cloud VM, NebuSec said. That combination turns routine web content into a path toward privileged tools and retained data.

NebuSec argued that agents that browse, call tools and store credentials test containers, VMs and kernels together, rather than as independent defenses. Its conclusion: “security controls can work as designed while still failing to prevent a complete attack chain.”

Meta's confirmation of the patches and any evidence of pre-patch data access will now determine whether the lab result becomes a broader security warning. Until then, the chain remains a contained demonstration with an unresolved real-user data question.

Source: epyhia.com

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,582 articles

Related articles

« Previous article