A critical zero-day vulnerability has been discovered in Meta's AI assistant Muse

In the architecture of the AI assistant Meta Muse a critical zero-day vulnerability has been identified, allowing third-party local applications to hijack control of the agent. The problem lies in the possibility of substituting the transcription endpoint through which the system processes incoming data.

Exploiting this vulnerability gives attackers full access to the user's account, despite Meta's declared security protocols. The attack is implemented by manipulating the configuration of the agent's interaction with system APIs, which allows malicious instructions to be embedded in the speech or text processing stream.

Researchers note that the architectural feature of Muse, which grants it privileged access to user data, makes the consequences of compromise critical. Unlike standard chatbots, the agent has the right to perform actions on behalf of the owner, which turns the vulnerability into a vector for stealing personal information and unauthorized account management.

Details of the technical analysis and discussion of the attack vector are available in the Hacker News community. At the moment, the company has not provided an official patch that eliminates the possibility of substituting the endpoint, which leaves users under threat of unauthorized access to their data.

The incident highlights the risks associated with integrating highly privileged AI agents into operating systems. The lack of strict isolation between local processes and the agent's control interface creates conditions for bypassing the authentication mechanisms built into the base model.