Security researchers at NebuSec said they identified a chain of vulnerabilities in Meta Muse, a personal AI assistant that operates through connected online accounts and runs each user session on a separate cloud-based machine. According to the researchers, the issues could allow a malicious webpage opened by the agent to access and modify private conversation history and stored memories without displaying a user approval prompt.
The demonstration was conducted in a controlled laboratory setting, NebuSec said. The researchers reported that the attack was limited to the targeted user’s virtual cloud environment and did not provide a path into the underlying shared infrastructure or other users’ machines.
Attack chain reached protected components
NebuSec’s account describes an initial web-based attack that could send Muse data to an attacker-controlled server. The researchers then reportedly combined a weakness in the assistant’s built-in browser with a Linux kernel vulnerability to obtain broader control of the user’s cloud machine.
That level of access could potentially expose files and credentials or tokens associated with linked accounts, according to the report. It could also affect the Sentinel monitoring layer, which is intended to control outbound activity and help isolate sensitive credentials from the AI agent itself.
Muse’s security architecture is designed around the assumption that an agent may encounter hostile content. While the reported controls contained the incident to a single cloud machine, NebuSec said the vulnerability chain could still bypass protections within that environment.
Fixes reportedly issued before disclosure
The firm said it privately disclosed the findings to Meta’s Muse team and that all identified weaknesses were corrected before the research was published. Meta had not publicly commented on the report or independently confirmed the remediation at the time of publication.
- The research was presented as a proof of concept rather than evidence of attacks against users.
- NebuSec did not report escaping from an individual cloud machine to shared infrastructure.
- The report did not state whether any user information had been accessed before the fixes were deployed.
The findings highlight the security challenges facing autonomous AI systems that browse the web, retain user context, and interact with connected services. Such systems can expand the impact of browser, operating system, and authorization flaws when weaknesses can be combined into a single attack path.
