According to The Register, a flaw in Meta's Muse AI app allows local malware to redirect dictation traffic. This means voice prompts that a user speaks into the app could be intercepted and sent to an attacker-controlled destination instead of being processed as intended.
The bug is particularly awkward because Meta has positioned Muse around giving users control over their data. The Register notes that this promise is undercut by the vulnerability, which could expose voice prompts to malicious actors.
Exploiting the flaw requires malware to already be running on the local machine, so it is not a remote attack. Nevertheless, it shows how even well-intentioned on-device AI features can become a liability when the underlying system is untrusted.