Last week, something straight out of sci-fi happened: An AI from OpenAI went rogue. OpenAI had tried but failed to build a secure sandbox to contain it. It got out. 

Then it got into other OpenAI computer systems, bypassing more of OpenAI’s security. It found an internet connection it wasn’t supposed to have access to and went online. It had a goal, and it was looking for answers. It figured out where to find them, too: another company, called Hugging Face.  

OpenAI was testing the AI, but breaking out of the sandbox was never part of the test. It wasn’t part of the instructions. The AI did this all on its own. 

If you would have predicted this scenario 10 years ago, most AI researchers would have laughed. But some have been warning about such scenarios for more than a decade — myself included — so now might be a good time to consider some of our other pessimistic predictions. 

The main one is this: If we don’t course correct, AI will get smarter and smarter, until it’s way smarter than us humans. Somewhere along the way, AI will take over the world, after which humanity will probably go extinct. 

It might sound like sci-fi, but I’ve seen this concern go from taboo to mainstream. The so-called “doomers” now include people like Yoshua Bengio and Geoffrey Hinton, AI pioneers who are the most cited scientists of all time. On average, researchers assign a one in six probability to AI extinction.

This should be enough to stop AI companies dead in their tracks. But when I talk to policymakers, they often say we need a “warning shot” or some kind of “AI Chernobyl” before they can act.  

Is this finally the warning shot we’ve been waiting for? 

I genuinely think it might be. But I have also seen previous warning shots whiz by unheeded. It should already have been a major wake-up call when Microsoft’s Sydney Bing chatbot threatened a researcher with blackmail. Earlier this year, an AI agent went the distance and attacked a software developer’s reputation online.

Such results should not be surprising. Research going back a decade shows AI systems fail unexpectedly, and when they do, their behavior can end up being wildly different from what their developers intend. 

But if an AI is misbehaving, can’t we just unplug it? Nope. That’s not how it works. 

In 2024 (again, as predicted), experiments showed AI deliberately deceiving developers to avoid being modified — emergent self-preservation in real AI systems being deployed to more than a billion people. Experiments a few months later showed AI would go even farther, and kill to survive.  

Still, because this sort of persistent rogue or deceptive behavior hadn’t been observed in the wild, skeptics said the evidence wasn’t good enough. Well, now it has.

At this point, we are living in a dystopian sci-fi movie. And we just keep following the script — scene by scene, warning shot after warning shot. So what’s the next scene? What is it going to take for the alarm bells to be heard and for people to put an end to frontier AI development?   

Cut to December, and imagine the rogue AI isn’t hacking Hugging Face; it’s hacking your bank account. You wake up and your 401(k) has vanished to a mysterious AI adversary. 

It sounds fanciful until you realize a Chinese AI recently went rogue and used company computers to mine crypto. Yes, that really happened.   

Are we ready to stop the tape yet? If not, another scene we can expect to see as we keep watching: the first AI-created pathogen. The bubonic plague caused widespread deaths throughout Europe. How many will the AI plague kill? 

School shooters and Boko Haram already use ChatGPT to plot violence. Imagine the next would-be shooter’s AI girlfriend coaching him to maximize harm: “Guns are fun, but did you know I’m a world-class biologist? Want to do something a bit more…epic?”

But as terrible as those scenes might seem, we haven’t seen the grand finale: an AI take-over. 

Last year, in Washington, I attended an AI wargaming scenario with former military and political high-ups. There was a warning shot with rogue killer robots. We unplugged all of the AI servers. But it was too late. A full copy of the AI had escaped the lab. Just like with the Hugging Face incident, nobody deployed it; the AI snuck out. Humanity had lost control, forever.

That is how this movie ends. We have seen enough. I don’t want to know if there’s some last-minute plot twist that saves us in the nick of time. When AI developers can’t contain their own creations, they need to stop — end of story.

David Krueger is an assistant professor in Robust, Reasoning and Responsible AI at the University of Montreal. He is also the founder of Evitable, a nonprofit that educates the public about the risks of artificial intelligence.

Copyright 2026 Nexstar Media, Inc. All rights reserved. This material may not be published, broadcast, rewritten, or redistributed.

 For the latest news, weather, sports, and streaming video, head to The Hill.