Loading SecureCyberVM…
SecureCyberVM
CVE Lookup Daily Threat Briefing Exploited CVEs Ransomware Zero-day
SecureCyberVM
Get the Patch Brief

Malicious webpage let researchers hijack Meta Muse agent and read private chats

NebuSec showed a malicious webpage could make Meta's Muse AI agent leak and alter private chats and memories, then escalate to root on the user's cloud VM. All exploits were patched before publication.

Malicious webpage let researchers hijack Meta Muse agent and read private chats

Researchers at NebuSec have shown that a single malicious webpage, opened by Meta’s Muse personal AI agent, could be used to read and alter a user’s private conversation history and stored memories without any approval prompt. By chaining a browser flaw with a Linux kernel vulnerability, the team went further and took full control of the cloud machine that runs an individual user’s agent, according to NebuSec’s write-up and reporting by The Hacker News.

NebuSec said it reported the findings to the Muse team and that every exploit described in its research had been patched before publication on November 20. No CVE identifiers have been assigned or disclosed for the browser or kernel issues, and Meta has not publicly commented on the research, according to The Hacker News.

What happened

Muse is Meta’s personal AI agent, launched on September 8, according to The Hacker News. It acts through online accounts that users connect to it and can make purchases on their behalf. Each user’s agent runs on its own dedicated cloud computer, and a monitoring layer called Sentinel is meant to gate outbound traffic from that environment.

NebuSec’s research, titled Hacking Muse: Stealing Private Chats Through Browser Use, describes a three-stage chain:

  • Browser foothold. A vulnerability in the Chromium-based browser that Muse uses for document tasks, such as creating PDFs, gave code execution inside the agent’s runtime container when the agent loaded attacker-controlled content.
  • Egress control gaps. Weaknesses in how Sentinel classified and inspected outbound requests allowed conversation history and memories to leave the environment without triggering an approval prompt. Stored memories and history could also be modified.
  • Container escape. A Linux kernel vulnerability in the shared kernel was used to break out of the container and gain root on the user’s virtual machine, exposing persistent storage, the secure credential store and Sentinel itself.

The researchers said their automated exploit generation framework produced kernel exploits that succeeded in more than 99.9% of test runs. Technical specifics of the browser and kernel flaws have not been tied to public identifiers.

Who is affected

The research concerns users of Meta Muse whose agents run in Meta’s cloud environment. According to NebuSec, all of the demonstrated exploits were fixed before the findings were published, so no user action is described as necessary to receive the fixes. The compromise in the demonstration stayed inside a single user’s virtual machine; the researchers did not claim access to the underlying host, the hypervisor or other users’ machines.

The Hacker News reported that the write-up does not say whether any real user data was accessed before the patches were deployed. Neither source describes the date the issues were reported or the date fixes shipped, and no bug bounty amount has been disclosed.

The research is separate from an earlier Muse issue disclosed in September, in which malware already present on a Mac could hijack the assistant on the device, according to The Hacker News.

Severity and exploitation

There is no indication of exploitation in the wild. The Hacker News characterized the work as a laboratory demonstration rather than an attack on real users. No CVSS score has been published by Meta or the researchers.

The potential impact, however, is significant for an agent designed to act on a user’s behalf. The chain reached the credential store and saved files on the user’s machine, and The Hacker News reported that the attack could be extended to Muse’s stored credentials for connected accounts. Neither source states that a specific credential was actually extracted during testing. The controls Meta built for Muse worked as designed, according to the reporting, yet the chain succeeded by targeting the layers beneath them.

What defenders should do now

  1. Inventory which AI agents in the organization can browse arbitrary web content and act through connected business accounts, including consumer agents employees may have linked to corporate services.
  2. Limit the scope of credentials and OAuth grants issued to agents, and prefer short-lived, narrowly scoped tokens that can be revoked centrally.
  3. Keep long-lived secrets outside the agent’s runtime where possible, as NebuSec recommends, so that a compromise of the execution environment does not expose them.
  4. Assess agent platforms as whole execution stacks, covering the browser, container, operating system kernel, network enforcement and credential services, rather than relying on a single policy layer.
  5. Review audit logs for connected accounts used by agents for unexpected actions, and establish a procedure to revoke agent access quickly if a vendor discloses a security issue.
  6. For internally built agents, patch headless browsers and host kernels on the same cadence as internet-facing servers, since agents routinely process untrusted content.

SCVM analysis

This case illustrates a pattern that defenders should expect to see more often: agent safety controls designed around prompt-level risks can be bypassed entirely by conventional memory-corruption and privilege-escalation bugs in the software stack that hosts the agent. A browser used for a mundane task such as PDF generation became the entry point, and a shared kernel became the boundary that failed.

The positive signals are that the issues were reported privately and fixed before disclosure, and that isolation held at the per-user virtual machine level. The open questions are transparency-related. Without CVE identifiers, a disclosure timeline or a statement from Meta, enterprises cannot easily judge how long the exposure window was or whether similar components are used elsewhere. Organizations adopting autonomous agents should treat them as privileged, internet-facing workloads and ask vendors for the same patch and incident disclosure commitments expected of any other critical SaaS provider.

Sources

About this article: produced by the SCVM Newsroom with AI assistance from the public sources linked above, enriched with NVD, FIRST EPSS and CISA KEV data, and checked by automated quality controls under our editorial policy. Spotted an error? Request a correction.

SecureCyberVM

Install SecureCyberVM

Get the Daily Threat Briefing and exploited-CVE alerts one tap away, full screen, like an app.