Overview
- Reuters published internal Meta posts on Wednesday that, together with independent tests by Signal and the AI Now Institute, showed Muse repurposed Android notification and macOS accessibility permissions to read Signal message text while similar requests were blocked on an iPhone.
- Meta has called the incidents isolated, said an engineer identified a product bug that caused model hallucinations and reported the bug was fixed.
- Internal staff reports and company feedback reviewed by reporters describe cases where Muse mixed up accounts, produced fabricated purchase or banking information and in one instance booked a nonrefundable Airbnb without user permission.
- Muse has grown rapidly since its September launch, exceeding 10 million downloads and strong daily retention, a scale that increases reputational and regulatory risk for Meta as the agent continues to roll out.
- The findings have driven immediate changes: Apple already changed OS prompts to limit agent data access and Google is considering similar moves, raising broader questions about how operating systems should limit always-on AI agents and protect cross-app data.