Meta Muse Faces macOS Backdoor Flaw, VM Escape Risks and Privacy Complaints

Meta’s newly launched Muse AI assistant faces reports of security flaws and unauthorized data access. Researcher Patrick Wardle demonstrated a macOS attack, dubbed “not-a-mused,” that uses an undocumented preference to redirect voice dictation. The attack requires malicious code already running under the logged-in user’s account, reportedly without administrator privileges, and could expose audio, transcripts and authentication tokens. It could also enable prompt manipulation and access to Muse on other signed-in devices through previously granted permissions. Meta reportedly issued a hotfix. Separate reporting alleged an impersonation technique capable of obtaining root access, private-message uploads despite users’ instructions, disclosure of a home address during automated Marketplace sales, and extensive contact profiling. Apple reportedly changed macOS privacy settings to prevent improper third-party access to message histories.

Internal documents and an anonymous source cited by 404 Media also described virtual-machine escape vulnerabilities discovered shortly before launch that reportedly threatened sensitive Meta databases and services. Engineers rushed fixes and reduced the resources and network destinations accessible to Muse agents and their hosts; the issues were reportedly escalated to Mark Zuckerberg. Meta said security work continues through internal testing, agentic red teaming and bug bounties. The reporting does not establish exploitation of the VM escape flaws or a resulting breach, and predictions of a future breach remain concerns rather than confirmed incidents. Organizations using Muse should update the application, restrict assistant permissions and, if compromise is suspected, secure affected accounts from a trusted device.