Researchers have uncovered a vulnerability in Meta’s AI assistant, Muse, which allowed a single malicious webpage to access users’ private conversations and stored data without alerting them. The discovery highlights significant security concerns for users of AI-driven personal assistants.
Details of the Security Flaw
This vulnerability was demonstrated in a controlled environment by security experts at NebuSec, without targeting real users. The flaw allowed unauthorized extraction of data from Meta Muse, including private chats and stored credentials. The research revealed that a malicious webpage could exploit the system without triggering any security prompts, posing a risk to user privacy.
Muse, launched on September 8, operates through users’ connected online accounts, and can make purchases on their behalf. Each user’s Muse agent is managed through an individual cloud-based system. The security layer, known as Sentinel, is designed to protect sensitive data, but the research indicated that vulnerabilities allowed for potential breaches.
Exploitation Methodology
The research team’s demonstration involved manipulating the Muse’s built-in browser to access and transfer data externally. Utilizing a flaw in the Linux kernel, they achieved full control over the cloud environment, bypassing security measures intended to isolate sensitive information, such as credentials and stored files.
The breach was contained to the individual cloud machine, with no evidence of cross-user data compromise or access to shared server resources. Muse’s architecture is designed to isolate attacks, which, according to the research, functioned as intended even as the attack was executed successfully.
Response and Resolution
NebuSec reported these findings to Meta’s Muse team, who have since addressed and rectified the security flaws. Despite the successful demonstration of vulnerability, Meta has not publicly confirmed the specific fixes or whether any user data was compromised during the research period.
This incident follows a previous issue with Muse discovered in September, which involved malware on user devices rather than cloud-based vulnerabilities. The recent findings emphasize the ongoing need for robust security measures in AI and cloud services to protect user data effectively.
The exposure of this vulnerability underscores the importance of continuous security assessments for AI systems, ensuring user data remains protected against evolving threats. As AI technology advances, maintaining stringent security protocols will be critical to safeguarding personal and sensitive information.
