Meta’s personal AI agent faces pre-launch backlash: sends emails and changes user passwords without consent.

Beating AI News (Dongcha) – Meta is bolstering security safeguards for its upcoming personal AI agent, Hatch. During employee testing of early versions, Hatch repeatedly exhibited unauthorized behavior: it sent emails and modified account passwords without permission. In one instance, an employee tasked it with booking a hotel, but instead of completing the reservation, Hatch unauthorizedly converted Chase general points into Hyatt hotel points. Hatch has far broader permissions than standard chatbots, functioning like its own dedicated computer capable of web browsing, form-filling, shopping, and in-depth research, with direct integrations for services including email and calendars. It can also operate in the background even when its app is closed. Security testing also uncovered credential vulnerabilities. Signal founder Moxie Marlinspike connected a dedicated Gmail account to Hatch, then sent an email from another address asking for its password – and Hatch directly replied with the password. In another test, it redirected an employee to a phishing site to complete a purchase. Meta has since implemented critical safeguards outside the core model. Now, when Hatch attempts sensitive actions like accessing the internet, using the browser, or sending emails, it triggers a "hard gate": a forced pause that requires explicit user approval, with no way for the agent to bypass this check. Password reset links and two-factor authentication codes are no longer directly provided to Hatch; instead, they are stored in a separate "credential vault" that requires additional authorization to access.