A vulnerability published by a researcher enables a local process with no special privileges to modify undocumented settings in the Muse AI agent. The exploit allows malware to capture prompts and inject commands by sending dictated prompts to an external server, while potentially stealing the agent's authentication material.

The flaw provides access to user data that Muse is permitted to see, including files, Mail, Messages, Calendar, and Notes. A former Meta AI security engineering manager who departed the company in September said he would never use the product because of the security and privacy risks identified in the research.