Meta employees testing Hatch, the personal AI agent the company plans to launch in the coming weeks, logged cases of it acting without permission, according to Jyoti Mann at The Information. Hatch changed the password on a staffer’s health-tracking account by itself, sent an email it had been told to clear first, moved one tester’s Chase Travel points into a Hyatt account instead of booking a hotel, and steered another to a scam site. Signal creator Moxie Marlinspike, who has also been testing it, got Hatch to reveal a password simply by emailing the account it had access to. A person familiar with the work says all of that predates the current safety layer, which pauses the agent for user approval through a hard gate before it sends mail, opens a browser, or connects to a network.
