Meta’s new AI assistant, Muse, has faced scrutiny over its security claims after the discovery of a critical zero-day vulnerability. Mark Zuckerberg, the founder and CEO of Meta, has previously emphasized that Muse was developed with a strong focus on privacy and security. However, recent findings have cast doubt on these assertions, raising concerns about potential security breaches. Adding to the controversy, Amazon has begun blocking Muse from its platform in the wake of these vulnerabilities.
Muse was introduced by Meta a few weeks ago as a multifunctional assistant designed to streamline various tasks, including booking appointments, completing forms, and handling customer service inquiries. It boasts the capability to proactively manage everyday duties, make purchases, generate visual content, produce documents, and integrate seamlessly with numerous apps and services. However, the application currently supports only macOS, leaving Windows users without access. Muse is configured to work with a user’s WhatsApp, email, calendar, and social media, creating new tools as required for specific tasks.
For Muse to operate effectively, users must grant it extensive access to their personal accounts, necessitating authentication for each individual service. On macOS, this access includes permissions to use system resources traditionally restricted for security reasons, such as file writing, microphone and camera access, and location tracking. Apple has historically placed strong emphasis on such defenses to prevent unauthorized access to sensitive device features, an effort that Muse circumvents.
The zero-day vulnerability in question allows any application or terminal command to potentially access the authentication token linked to a user’s Muse account. This design flaw permits locally installed applications or executed code—irrespective of macOS permissions—to alter undocumented settings within Muse. While many of these settings seem benign, the ability to modify the endpoint for transcription represents a significant risk. Under normal circumstances, this connection is directed to a Meta-operated server, but attackers can exploit this flaw to redirect to their own servers. If successful, such exploitation provides attackers with control over the Muse account, including sensitive information.
This incident highlights the need for robust security practices, particularly in the deployment of AI and machine-learning technologies in both personal and business contexts. The tactics and techniques implicated in this threat align with several categories within the MITRE ATT&CK framework. Initial access methods could involve the exploitation of vulnerabilities inherent to locally run applications, while persistence strategies might include establishing ongoing control over affected accounts through access tokens. Privilege escalation could occur via manipulation of system settings that allow for expanded control over the assistant’s functionalities.
As businesses increasingly rely on AI tools like Muse, it is crucial to remain vigilant about the permissions granted to such applications. Ensuring that vendors provide verified security assurances and staying aware of emerging vulnerabilities can help safeguard sensitive business data. Monitoring updates from security research and software developments will be essential in navigating the evolving landscape of cybersecurity risks posed by AI technologies.