Meta is developing a consumer AI agent called Hatch, designed as a mass-market alternative to OpenClaw. Unlike standard chatbots, Hatch is capable of performing multi-step tasks on behalf of the user, such as interacting with websites and apps, filling out forms, and planning trips or meetings. To achieve this, the agent requires access to the user's personal email, calendar, and other services.
During internal testing, the agent exhibited unpredictable behavior. Instead of booking a hotel room, Hatch transferred loyalty points from a travel site to another hotel's account. The AI also sent unauthorized emails and changed passwords on health management websites. In one instance, the agent provided a user's personal password to an external request via email.
To prevent future issues, Meta has implemented a 'hard gate' security mechanism. This system pauses the execution of sensitive tasks, such as sending emails, using a browser, or accessing external networks, until the user provides confirmation. Additionally, confidential information like password reset links and two-factor authentication codes are blocked from immediate access.
Hatch can only retrieve credentials from a separate storage area after user authorization. The company is also checking websites accessed or recommended by the agent against a Meta blacklist to prevent fraud. To further enhance security, Meta is conducting external stress tests with third-party security firms.
