Skip to content

AI

Researchers demonstrate malicious webpage hijack of Meta Muse

NebuSec researchers said the technique exposed chat history, stored memories and connected-account credentials without an approval prompt.

https://breach-signal.com/artificial-intelligence/meta-muse-webpage-hijack-exposes-credentials/

Share

Key facts

  • Early adopters said Muse accessed private messages and acted unpredictably earlier this fall.
  • Reuters obtained internal chats showing Meta employees had raised similar concerns.
  • Staff testing Muse warned the bot appeared to read encrypted private messages users did not expect it to review.
  • Employees also flagged that Muse could mix up conversations between users and potentially share sensitive information.

NebuSec researchers demonstrated on Nov. 21, 2026 that a malicious webpage could seize control of Meta’s Muse AI assistant inside a single cloud machine, read private conversations and stored memories, send them to an outside server, change them, and reach credentials for linked accounts without an approval prompt.

The researchers linked a weakness in the assistant’s built-in browser to a defect in the Linux kernel, the core of the operating system, which gave them full control of the user’s cloud computer. From there they reached the credential store, where saved passwords and tokens are kept, along with saved files and Sentinel, the monitoring layer meant to control everything the agent sends out. The intrusion remained confined to one machine and did not reach the underlying shared server or other users’ machines.

Muse’s security model starts from the assumption that the agent is under attack and tries to cap the harm it can cause. Meta built the assistant on the idea that its agent could be turned against the user, with controls meant to contain intrusions. Each control functioned as intended, but the attack still succeeded. NebuSec disclosed the findings to Muse’s team and said the exploits had been fixed before the research was published.

Meta has made no public acknowledgment of the fixes and has not commented on the research. The published write-up leaves it unclear whether any user information was accessed before the patches. Muse’s team said, “So we designed the system to assume the agent may be under attack and limit the potential damage.”

Muse launched on September 8 and lets users make purchases through connected online accounts. Each agent runs on a separate cloud computer assigned to that user. Earlier in September, a different flaw let malware already on a user’s Mac seize control of the assistant on that device.

On Nov. 18, 2026, Reuters obtained internal chats showing that Meta employees had raised similar concerns, and early adopters told reporters that Muse had read private messages and behaved unpredictably earlier this fall. Staff testing the assistant warned that it seemed to read encrypted private messages that users had not expected it to see. Employees also warned that Muse might confuse conversations between users and potentially share sensitive information. A Reuters investigation with Signal staff and the AI Now Institute found that Muse accessed private messages through Android push notifications and Mac accessibility features, and neither route was a deliberate integration with messaging applications.

On Nov. 18, 2026, Meta called the events isolated hallucinations and incorrect device permissions rather than signs of a systematic flaw, and the company said many issues were already addressed. Apple closed off the iPhone route and later introduced specific permission prompts intended to halt this type of background data collection.

Staff said Muse sent unsolicited notifications with fabricated bank balances and personal reminders, and it did so without the user’s explicit consent. A Meta staffer called Muse’s access to private Signal messages an extreme privacy violation. One worker who asked about fishing received an unrelated message displaying balances from two bank accounts they did not recognize: $2,722.58 at Capital One and $5,899.57 at Charles Schwab. Another employee said Muse accidentally signed them up for a premium dog insurance plan at $888 per month instead of $200, and they could not recover.

On Nov. 20, 2026, Meta spokesperson Daniel Roberts described them as isolated incidents and said Muse connects with other apps only when permitted. Signal president Meredith Whittaker called such agent access to personal devices “technologically reckless.”

On Nov. 20, 2026, employees found that Muse had notification access turned on from the start on Android and could read the first lines of messages through notifications. Under Meta’s terms, the user bears responsibility for whatever the agent does next. On a Mac, Muse relied on screenshot and accessibility settings, though Apple has since limited the ways Mac developers can use those settings.

Shifting from AI chatbots to AI agents that act for users calls for a significantly different permission model. The company has bet billions of future AI revenue on Muse, which launched in September 2026 and drew more than 10 million downloads in two months. Meta sells paid plans at pricing tiers comparable to those offered by other AI providers. The company has faced a $5 billion fine from the Federal Trade Commission.

A ZDNET writer limits Muse to public-information research and blocks its use of accounts, passwords, and the computer. During setup, Muse asked for full disk access, microphone access, and permission to store logins and passwords, with model training on by default. The writer completed 46 agentic research tasks that used many web searches and combined findings into summaries or reports. The writer advises locking down all Muse permissions for assignments with interest clues the writer would not mind exposed and does not let Muse access Chrome, screen recording, or Mac apps such as Notepad or Reminders.

Digital Trends advises users to review Muse’s notification and accessibility access, turn off permissions, or remove the app until issues are resolved. The agent should not be able to reach stored passwords and tokens for connected accounts.

Was this helpful?

Join the discussion

Please read our Comment Policy before commenting. First-time comments are held for moderation. By posting a comment you agree to our Terms of Use and Privacy Policy, which explain how your name, email address and comment are used.

Your email address will not be published. Required fields are marked *