Security

One Malicious Webpage Let Researchers Hijack Meta’s Muse Agent and Read Private Chats

Leave a comment

Security firm NebuSec published research on November 20 showing that a malicious webpage could make Meta’s Muse agent hand over a user’s private conversations and stored memories, with no approval prompt, The Hacker News reports. Muse is Meta’s personal AI agent, launched September 8, that works through accounts you connect and can make purchases for you. Each user’s agent runs on its own cloud computer, with a monitoring layer called Sentinel meant to gate everything it sends out.

In the lab demo, the page pulled chat history and memories to an outside server and altered them. The researchers then chained a flaw in Muse’s built-in browser with a Linux kernel flaw to take full control of the user’s cloud machine, reaching the credential store, saved files and Sentinel itself. It didn’t escape to the shared server or other users. This was a demonstration, not an attack on real users, and NebuSec says Muse’s team fixed everything before publication. Meta hasn’t publicly commented.

My take

Muse was designed to assume its own agent could be turned against you, and every control “worked as designed,” yet the chain still went end to end. That’s the lesson for all of us running agents: a browsing agent reading untrusted pages is a prompt-injection target, so keep its credentials and its ability to act as far apart as you can.

God Bless, and Happy hacking!

References & Further Reading

Subscribe (RSS) Follow Security via RSS

Leave a note

Your email address will not be published.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

This site uses Akismet to reduce spam. Learn how your comment data is processed.