Researchers show webpage attack on Meta Muse cloud agents

Researchers show webpage attack on Meta Muse cloud agents

Add CyberWorldOpsto your preferred sources on Google
Researchers show webpage attack on Meta Muse cloud agents
AI

Illustrative image generated with AI

A malicious webpage could steal and alter a Meta Muse user’s private conversations and stored memories without an approval prompt, security firm NebuSec demonstrated in laboratory research, The Hacker News reported. The researchers also chained weaknesses to take full control of that user’s cloud computer, reaching credentials for connected accounts. NebuSec says Muse’s team fixed every weakness before publication.

The findings concern Meta’s personal AI agent, launched on September 8, which works through online accounts that users connect and can make purchases on their behalf. Each user’s agent runs on a separate cloud computer. The demonstration involved that cloud environment and was not an attack on real users.

The attack remained inside one user’s machine: the researchers did not reach the shared server underneath it or other users’ computers.

From a webpage to the credential store

Muse’s security design anticipates that an attacker might turn the assistant against its user. A monitoring layer called Sentinel is intended to act as the single checkpoint for anything the agent sends out, while passwords and tokens—credentials that let software access connected accounts—are meant to remain beyond the agent’s reach.

The demonstration began with a malicious page opened by the agent. The page could cause conversation history and stored memories to be sent to an external server and could alter that information, without showing the user a permission request.

The researchers then combined a flaw in Muse’s built-in browser with a Linux kernel flaw, a weakness in the operating system’s core. Muse uses the browser for tasks such as creating PDFs. Chaining the two flaws gave the researchers full control of the user’s cloud computer, as described in their technical breakdown.

That control extended to saved files, the credential store and Sentinel itself—components that Muse’s design separates from the agent. NebuSec said the individual containment controls worked as designed, yet the complete attack still succeeded within the user’s machine.

The cloud compromise differs from an earlier Muse flaw found in September, which allowed malware already running on a user’s Mac to hijack the assistant locally.

Researchers say fixes are complete

NebuSec reported the findings to Muse’s team. Its account leaves any access to user data before patching unresolved, and Meta had not publicly confirmed the fixes or commented on the research when The Hacker News reported it.

For users with connected accounts, the latest reported development is NebuSec’s statement that Muse’s team fixed all the described exploits before the research became public.

Read next

Sources

Back to home

Latest Cybersecurity News

All cybersecurity news →