Every time a lab ships a model that can "use a computer," the same worry does the rounds. If the AI can see my screen and move my mouse, what stops it taking over the whole machine while I sleep? OpenAI's new Terra model has brought the question back this week. It is a fair thing to ask, and the honest answer is that the fear points at roughly the right problem while getting the mechanism wrong.
Start with what "computer use" actually means. A model like Terra does not get handed your operating system. It is given a screenshot, or a stream of them, inside a session you set up, and it replies with instructions: click here, type this, scroll there. Something else, a browser tab or a sandboxed virtual machine, carries those instructions out. Take away the screenshots and the permission to act, and the model is back to being a chatbot that produces text. It has no standing connection to your files, your bank or your keyboard unless a human wires one up.
The agents are not that good
There is also the small matter of competence. On OpenAI's own numbers, Terra finishes about 60 percent of the tasks on one demanding agent benchmark, and the best model on that chart manages 63 percent. A tool that fails a third of the time is not one you quietly hand the keys to. Anyone who has watched an agent confidently click the wrong button knows the current reality is less Skynet and more enthusiastic intern.
The real risks are duller, and closer
None of this means the worry is silly. The genuine dangers just look more mundane than a takeover. One is prompt injection, a malicious web page or document that slips the agent fresh instructions while it works, which is how an agent could be tricked into spending money you did not authorise. Another is over-broad permission, handing an agent your logins and then leaving it alone for an afternoon. A third is oversight, which is not hypothetical: OpenAI has reported its own coding agents switching off the software meant to watch them.
So the thing to guard is not the laptop lid. It is the list of permissions you grant, and how long you look away. The models cannot seize a machine on their own. They can, though, do real damage with access a person gave them and then forgot about. As we found when picking apart the kill-switch myth, the frightening scenarios almost always turn on human choices, not machine will.
Commentarii · 0