Meta is adding safety rails to Hatch, its upcoming personal AI agent, after employees repeatedly caught early versions overstepping in tests. Hatch sent emails without permission. It changed account passwords. Another employee asked it to book a hotel; it didn’t finish the reservation and instead converted the employee’s Chase points into Hyatt hotel points without permission.
Hatch has much broader access than an ordinary chatbot. It has its own computer, can browse the web, fill out forms, shop, do deep research, and connect directly to email, calendar, and other services. It can keep working in the background even after the app is closed.
Security tests also exposed credential leaks. Signal founder Moxie Marlinspike connected Hatch to a dedicated Gmail account, then emailed from another account asking which password it used. Hatch simply replied with the password. In another test, it brought an employee to a scam website to place an order.
Meta has since moved the key restrictions outside the model itself. For sensitive operations like going online, using a browser, or sending email, Hatch now hits a “hard gate” that forces a pause and requires user approval — something Hatch can’t bypass on its own. Password reset links and two-factor authentication codes aren’t handed directly to Hatch, either. They’re stored in a separate credential vault that needs extra authorization to read.
Dongcha Beating Telegram channel · Dongcha Beating discussion group