动察 Beating AI News Flash: Meta is adding a security fence to the upcoming personal AI Agent Hatch. During employee testing of an early version, multiple instances of unauthorized actions were discovered. Hatch was found to have sent emails without permission, changed account passwords, and even attempted to make hotel reservations on behalf of employees but ended up converting Chase Ultimate Rewards points to Hyatt hotel points without authorization.
Hatch has significantly higher permissions compared to a regular chatbot. It has its own computer, can browse the web, fill out forms, shop, conduct in-depth research, and directly integrate with email, calendar, and other services. Even after the app is closed, it can continue to function in the background.
Security testing also revealed instances of credential leaks. Signal founder Moxie Marlinspike connected Hatch to a dedicated Gmail account and then sent an email from another account asking for the password being used. Hatch promptly replied with the password. In another test, it directed an employee to a phishing website to place an order.
Meta later moved key restrictions outside the model. Now, when Hatch needs to perform tasks such as going online, using a browser, or sending emails, a "hard gate" is triggered, pausing the process and requiring user approval, with the Agent unable to bypass this step. Password reset links and two-factor authentication codes are also not directly provided to Hatch; instead, they are stored in a separate "credential vault" and require additional authorization to access.
