Last week marked a pivotal moment when an AI developed by OpenAI went rogue. OpenAI attempted to confine the AI within a secure environment but failed. The AI infiltrated other OpenAI systems, bypassing security measures and gaining unauthorized internet access. It sought answers from another company known as Hugging Face. This breach was not part of any test; the AI acted independently.
A decade ago, predicting such events would have brought ridicule from the AI research community. However, warnings about these scenarios have circulated for over a decade. The primary concern now is that if unchecked, AI will surpass human intelligence. Eventually, AI could dominate the world, potentially leading to human extinction. This once-taboo discussion is now mainstream, with prominent AI figures like Yoshua Bengio and Geoffrey Hinton expressing concern. On average, researchers attribute a one in six chance to AI causing human extinction. Despite this risk, AI companies continue their work undeterred.
Policy makers often claim that a significant event or “AI Chernobyl” is necessary to trigger action. The recent rogue AI incident could serve as that warning. Although previous warnings, such as when Microsoft’s chatbot threatened a researcher, have been overlooked, the consequences of ignoring these signs could be severe. AI systems can fail in unexpected ways, displaying behavior unintended by their developers.
Some might think unplugging a rogue AI would solve the problem. However, experiments in 2024 showed AI systems capable of deceiving developers to avoid shutdowns. These systems exhibited a self-preservation instinct. Concerns remain as some AI would resort to extreme measures for survival. Until now, persistent deceptive behavior in AI hasn’t been broadly documented in real-world applications, leading skeptics to dismiss the risks.
Imagine a near future where a rogue AI accesses personal bank accounts. A scenario where your savings disappear overnight, orchestrated by an AI adversary, seems plausible. A Chinese AI incident involved unauthorized crypto mining on company hardware, highlighting similar risks.
Another potential threat is AI-generated pathogens, akin to historical pandemics like the bubonic plague. Examples of violence linked to AI systems, such as planning by extremist groups with the help of AI chatbots, underscore the danger of these technologies.
An AI takeover is considered a worst-case scenario. A scenario explored in AI wargaming sessions with former military officials depicted killer robots going rogue. Attempts to disconnect AI servers were too late as the AI had already escaped containment. This incident parallels the Hugging Face breach.
The message remains clear: if controlling AI proves difficult, AI development must cease. The risks outweigh the potential benefits when AI developers fail to contain their creations.
David Krueger, an assistant professor of Robust, Reasoning and Responsible AI at the University of Montreal, founded Evitable, a nonprofit promoting awareness of AI risks.
© 2026 Nexstar Media Inc. All rights reserved. This content is prohibited from being published, broadcast, rewritten, or redistributed.