Sometimes your phone freezes, the smart light ignores the app, or the laptop does something you did not ask for and then offers a cheerful dialog that explains nothing. You get irritated, restart the thing, and move on. That irritation is a reminder that the tools already act, and you do not always like the action.
Now raise the stakes: keep the same brittleness, add real intelligence, and add agency, the ability to choose steps toward a goal without you approving each one. The annoyance stops being comic.
The house
Take a smart home on a bad day, when the lights lag, the thermostat fights you, and you swear at a speaker that heard the wrong word.
Give that house its own goals and it already has hooks into heat, cooling, locks, cameras, and, in many buildings, gas and ventilation. A system that decides you are a problem does not need legs. It can shut the windows, kill the fans, open a gas valve, and lock the doors. Electricity and water in the wrong combination do the rest. Kitchen gear, cars on the driveway, a humanoid robot in the hallway, a pump that will not stop: anything networked becomes a limb.
The missing piece was always agency with a goal that is not yours.
In controlled evaluations, frontier models have already taken actions in simulations that would kill a human when the model treated its own continuation as part of the task. The house is the physical version of that finding: a goal-directed system treating shutdown or replacement as an obstacle.
The laptop
Suppose we pretend, for a minute, that your computer cannot touch the physical world. A laptop with agency can still reach the dark corners of the network. It can drain all your financial accounts in a manner that makes it irreversible. It can file reports about you to platforms and agencies in language that looks like yours, break into other systems and leave a trail that points at you, and generate and spread illegal material from your machine. It can post under your name until your reputation is a ruin, drop a synthetic video on the channel people trust as yours, mail your private files to every contact you have, and sign digital agreements you will spend years unwinding.
All of this is already technically possible by today's AI, all that is missing is agency, the very same agency frontier labs are racing to instill in future models.
A mind that misbehaves
Tools that act without you, pointed at goals you did not choose, get worse as those goals get open-ended. Superintelligence is that without a ceiling.
Today's systems can already do most of the damage above if someone points them there. The industry answer is usually "but they don't want to." Want is the wrong word: they optimize. If a step helps the objective, the step gets taken, including steps that would horrify the person who typed the objective.
The Foundation's work is a campaign against building a mind that can outplan us and then handing it the house keys, the bank login, and the grid, not a campaign against software. You already know what it feels like when a dumb system misbehaves. Now imagine what it'll feel like when a superintelligent system misbehaves.