Meta Muse AI app flaw lets local malware redirect dictation traffic

Meta's Muse AI assistant app, touted for its security, may contain a vulnerability that allows local malware to redirect dictation traffic and potentially access sensitive information. The app utilizes a feature called Muse Secure VM, which grants users control over its access levels. However, security researcher Patrick Wardle discovered what he calls a "local zero-day" that could enable an unprivileged local process to modify Muse's settings and redirect dictation traffic to an attacker-controlled endpoint.

This flaw, if exploited, could lead to prompt injection, theft of authentication material, and abuse of user-granted access. Wardle likens the situation to living in an apartment building, where a malicious neighbor could gain access to all apartments. While Apple has implemented strong security measures like its Transparency, Consent, and Control (TCC) framework, Wardle worries that AI apps, due to their extensive access requirements, could become a single point of failure and undermine operating system security controls.

He also questions whether AI companies are prioritizing their own app security. Wardle suggests that Apple's on-device local dictation service could mitigate the risk, but Meta opted not to use it, potentially increasing the attack surface due to their desire for user data. As AI companies race to develop new technologies, Wardle believes privacy and security are not top priorities. Meta declined to comment on the vulnerability.