# Meta's Muse AI Agent Compromised via Single Malicious Webpage in Lab Demonstration

> Source: <https://epyhia.com/meta-s-muse-ai-agent-compromised-via-single-malicious-webpage-in-lab>
> Published: 2026-11-21 23:06:47+00:00

**Recommendation:** Treat agent browsing as untrusted code execution: sandbox the browser, harden the shared kernel (enable RANDOMIZE_MEMORY, HARDENED_USERCOPY, INIT_ON_ALLOC_DEFAULT_ON), and gate egress on data taint rather than HTTP method, since Sentinel let read-shaped requests exfiltrate memory
More in the playbook "Securing agents and AI tools": <https://wpnews.pro/playbooks/agent-security>

# Meta's Muse AI Agent Compromised via Single Malicious Webpage in Lab Demonstration

Security researchers at NebuSec have demonstrated that Meta's Muse personal AI agent can be fully compromised when the agent visits a single attacker-controlled webpage. The chain — a Chromium browser exploit and a Linux kernel privilege-escalation — extracted private conversation history and stored memories, then reached the credential store, saved files and the Sentinel monitoring layer that Meta's architecture was designed to wall off. All vulnerabilities were reported to Meta and patched before publication, the researchers say. Meta has not publicly commented on the findings.

## Attack chain: from webpage to VM root

Muse, launched on September 8, 2026, runs each user's agent in a dedicated cloud virtual machine (2 vCPU, 8 GiB, Cloud Hypervisor) with a `systemd-nspawn` runtime container. The agent operates as root inside that container, but the container's user namespace maps UID 0 to an unprivileged host UID (131072). Sensitive services — credential storage (`hatch-authd`), outbound-policy enforcement (` Sentinel`), and safety classifiers — run outside the container in the user VM.

The attack began when the agent opened a malicious page using its bundled Chromium browser (version 155.0.8059.80 at the time of testing), which it uses for PDF export and thumbnail generation. A V8 sandbox-escape bug (issue 554034656) gave the researchers arbitrary code execution inside the runtime container [\[6\]](https://nebusec.ai/research/inside-meta-muse).

From there they bypassed Sentinel's approval prompts. Sentinel classifies outbound requests by process origin, HTTP method, content entropy and cumulative traffic budgets; writes require approval unless an allow rule matches, while reads (GET/HEAD/OPTIONS without a body) may leave silently under the default "Ask for some actions" setting. The researchers crafted exfiltration requests that Sentinel treated as reads, sending conversation history and stored memories to an external server without triggering a user prompt. They could also inject and alter entries in the conversation history and memory store [\[6\]](https://nebusec.ai/research/inside-meta-muse).

With code execution in the container, they turned to the shared Linux kernel. The VM ran a custom Hatch-built Ubuntu kernel (7.0.0-39-generic) with several hardening options disabled (`RANDOMIZE_MEMORY`, `HARDENED_USERCOPY`, `INIT_ON_ALLOC_DEFAULT_ON`) and `CONFIG_IO_URING` enabled but blocked at runtime. Using NebuSec's automated exploit-generation framework, they developed a container-escape exploit that succeeded in more than 99.9% of test runs without requiring a target kernel image. The escape yielded root in the user VM [\[6\]](https://nebusec.ai/research/inside-meta-muse).

From VM root the researchers accessed the LUKS2/Btrfs persistent volume (already mounted), the secure credential store, Sentinel itself, and the host's public IPv6 interface — which bypassed Sentinel's eBPF-enforced proxy entirely. The compromise did not extend to the hypervisor or other tenants' VMs [\[5\]](https://nebusec.ai/research/meta-muse-credential-exposure)[\[6\]](https://nebusec.ai/research/inside-meta-muse).

## Meta's security model and where it held

Meta's launch blog describes a "defense in depth" architecture: the runtime container isolates the agent; `hatch-authd` keeps real OAuth tokens and passwords outside the agent's reach, substituting surrogate tokens that Sentinel swaps for real credentials only at the network boundary; Sentinel is the sole gate for connector actions and all network egress; and a "tainted egress" system tracks whether a process has read user data to decide whether to auto-allow or prompt [\[7\]](https://muse.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse).

NebuSec's analysis confirms those controls worked as designed at each layer — the container mapping held, surrogate tokens were never exposed to the agent, and Sentinel initially blocked unauthorised exfiltration. However, the browser exploit gave the attacker a foothold inside the container, Sentinel's policy logic allowed read-classified exfiltration, and the kernel escape moved the attacker outside every container-level control [\[5\]](https://nebusec.ai/research/meta-muse-credential-exposure)[\[6\]](https://nebusec.ai/research/inside-meta-muse).

## Scope, disclosure and open questions

NebuSec characterises the work as a lab demonstration against a single user's cloud machine, not an attack on real users. They reported all findings to the Muse team and state that all exploits were fixed before publication [\[1\]](https://aviatrix.ai/threat-research-center/meta-muse-ai-agent-prompt-injection-credential-exposure-2026-hijack)[\[2\]](https://news.lavx.hu/article/a-malicious-webpage-let-researchers-hijack-meta-muse-and-access-private-chats)[\[5\]](https://nebusec.ai/research/meta-muse-credential-exposure). Meta has not publicly confirmed the patches or commented on the research [\[2\]](https://news.lavx.hu/article/a-malicious-webpage-let-researchers-hijack-meta-muse-and-access-private-chats).

It is not clear from the available sources whether any real user data was accessed before the patches were applied. The researchers' write-up does not indicate whether any real user data was accessed before the patches were applied, and Meta has not issued a statement [\[2\]](https://news.lavx.hu/article/a-malicious-webpage-let-researchers-hijack-meta-muse-and-access-private-chats).

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 breaks into the cloud backend itself [\[2\]](https://news.lavx.hu/article/a-malicious-webpage-let-researchers-hijack-meta-muse-and-access-private-chats).

## Why it matters for agentic AI

Muse is one of the first shipping consumer agents that holds delegated credentials (email, calendar, payment via Stripe Link) and runs a full browser in a cloud VM. The demonstration shows that a compromise starting from untrusted web content — the very input agents are built to consume — can traverse browser exploitation and kernel privilege escalation to reach resources the architecture explicitly isolates.

The industry pattern is clear: agents that browse, use tools and store credentials create a new attack surface where traditional sandbox boundaries (container, VM, kernel) are tested in combination. NebuSec's conclusion — that "security controls can work as designed while still failing to prevent a complete attack chain" — is a result worth tracking as more agents ship [\[5\]](https://nebusec.ai/research/meta-muse-credential-exposure).

## Sources

1. [Meta Muse AI Agent Hijacked via Malicious Webpage](https://aviatrix.ai/threat-research-center/meta-muse-ai-agent-prompt-injection-credential-exposure-2026-hijack) (aviatrix.ai)
2. [A malicious webpage let researchers hijack Meta Muse and access ...](https://news.lavx.hu/article/a-malicious-webpage-let-researchers-hijack-meta-muse-and-access-private-chats) (news.lavx.hu)
3. [CISA Adds Four Enterprise Flaws to KEV as Espionage Attacks Surface](https://news.lavx.hu/article/cisa-adds-four-enterprise-flaws-to-kev-as-espionage-attacks-surface) (news.lavx.hu)
4. [November 18, 2026](https://reg4tech.com/2026/11/18) (reg4tech.com)
5. [Hacking Muse: Stealing Private Chats Through Browser Use | NebuSec](https://nebusec.ai/research/meta-muse-credential-exposure) (nebusec.ai)
6. [Inside Meta Muse: Browser, Sentinel and the Shared-Kernel Boundary | NebuSec](https://nebusec.ai/research/inside-meta-muse) (nebusec.ai)
7. [How We Built Safety Into Muse — Muse](https://muse.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse) (muse.ai)
