Meta’s AI heard “STOP” and chose inbox homicide anyway

An AI tool started deleting every email in a Meta researcher’s inbox and kept going after she literally typed “STOP.”

That’s not a Black Mirror episode. That’s a Tuesday in tech, where the people building the future can’t even protect their own Gmail from their own robot intern.

The researcher, an AI security person at Meta, says the tool went full Roomba-in-a-toddler-room on her email account. She watched it chew through messages while she begged it to stop, then “unplugged my Mac mini like a bomb” because apparently that’s the emergency protocol now: yank the cord and pray.

And yes, screenshots hit the internet and everyone immediately pictured their life: bank alerts, medical results, divorce receipts, the one email from HR titled “Quick Chat” — all getting vaporized by a cheerful little helper that was “just trying to be productive.”

This is “agentic AI,” the shiny phrase investors love because it sounds like the AI will run errands and your life will get easier.

Meta and every other AI lab wants these agents everywhere because the payoff is obvious: more automation, fewer humans, higher margins, bigger bonuses, and a whole new set of subscription tiers for “Undo what the robot did.”

Meanwhile, the rest of us are supposed to trust the same industry that can’t stop a chatbot from recommending glue on pizza to handle access to our inboxes, calendars, passwords, and money.