Meta's AI agent Muse had a hole big enough to hand your account to whoever got code running on your Mac.
Security researcher Patrick Wardle found a zero-day in the Muse macOS app that let an attacker with local access redirect the app's transcription processing away from Meta's servers and onto one they controlled. Ars Technica reports that hijack gave the attacker access to the victim's Muse account. The flaw traced back to an undocumented setting that any app on the machine could reach and modify. Meta has since shipped a patch.
Two design choices made this possible: Muse processes dictation in the cloud rather than on-device, and it left its internal settings open to any local process rather than locking them down. Cloud processing is a common tradeoff for AI features that need heavier compute than a laptop can offer, but it only works if the handoff to the server is authenticated and locked to the app itself. Here it wasn't.
The bug required local access, so this was never a remote, click-a-link kind of threat. But "local access" covers a lot of ground - a compromised app, a malicious browser extension, or physical access to an unlocked laptop would all qualify. For an AI agent that's meant to act on a user's behalf, that's a fairly low bar for someone to end up wearing your account.