Meta launches AI agent that can access other apps to send emails, make payments

Key Points:

• Meta is launching its Muse AI agent, internally called Hatch, aiming to provide "personal superintelligence" to billions of users, initially available in the U.S. via a dedicated app and WhatsApp, with plans to integrate it into smart glasses.
• The basic Muse agent will be free, with subscription tiers at $20 and $100 per month for heavier use, and users can opt out of having their interactions used to train Meta's AI models; an encrypted version is expected later this year.
• Muse can access multiple personal apps like email, calendar, payments, health, and smart home devices, running on its own cloud-based virtual machine to handle requests even when users are inactive, raising significant safety and privacy concerns.
• Meta delayed Muse's release to enhance security, implementing safeguards including a monitoring agent that can require authorization for certain actions, though internal testing has revealed mixed results, with some employees reporting useful applications and others encountering serious security flaws and reliability issues.
• The launch occurs amid broader industry challenges with AI agents causing unpredictable behavior, with Meta experiencing a 40% increase in technical and security incidents linked to AI development, resulting in a 70% rise in staff time spent managing these issues.