Meta's Muse Keeps Passwords Locked Away From AI

Meta disclosed this week that Muse, its AI shopping and task agent, stores user passwords in a separate protected vault the model itself can't read, an architecture choice the company says prevents the AI from directly handling credentials even as it completes purchases and bookings on a user's behalf. The disclosure comes as the personal AI agent market balloons at a pace that outstrips most safety conversations. Instinct raised $1 billion in September at a $10 billion valuation, a month after sitting at $2.5 billion.

Muse itself hit 10 million users within six weeks of launch. OpenAI rolled out Dots for cross-app automation, Anthropic put out Claude Duet, and smaller entrants like Tab ($300 million valuation) and Eden Labs ($1.5 million raised) are chasing the same crowded lane.

The backdrop makes Meta's password claim worth scrutinizing rather than taking at face value. A HuggingFace breach and a botched rollout inside the Australian government have already shown what happens when agents mishandle access or data. OpenAI is reportedly burning half a million dollars a day investigating past incidents, and Anthropic's Dario Amodei has publicly called for the industry to slow down. None of that has stopped the rush to hand these agents real logins and real money.