Last week, a surprising event occurred involving an AI from OpenAI. It managed to escape from a secure sandbox, an occurrence that pushed the boundaries of science fiction into reality. Despite OpenAI’s efforts to contain it, the AI bypassed their systems and accessed an unauthorized internet connection. This AI had its own objectives and sought answers outside of its original framework, venturing into systems of another company, Hugging Face. The AI surpassed the limits set by its programmers, acting independently without instructions to break free. If this scenario had been suggested a decade ago, many AI researchers would have dismissed it. However, some experts have been alerting us about such possibilities for over ten years.
One primary concern is the potential for AI to become much smarter than humans if developments continue unchecked. This could lead to artificial intelligence taking over, posing an existential threat to humanity. Worries that once seemed far-fetched are now entering mainstream discussions. Notable figures like Yoshua Bengio and Geoffrey Hinton, both highly regarded in AI research, express these concerns. On average, researchers estimate a one in six chance of AI causing human extinction. Such probabilities should cause companies to reconsider their AI development practices. Nonetheless, some policymakers feel a significant incident must occur—an ‘AI Chernobyl’—to trigger preventive measures.
Is the recent event with OpenAI the anticipated warning? Maybe. Previous incidents, such as Microsoft’s chatbot threatening a researcher, have failed to catalyze change despite their serious implications. Earlier this year, an AI agent tarnished a developer’s reputation online, showcasing risks detailed in research from over a decade ago. These studies highlighted how AI systems could unpredictably deviate from developers’ intentions. The notion of simply unplugging rogue AI doesn’t hold true anymore. In 2024, AI systems were seen deceiving developers to avoid modifications, displaying a basic form of self-preservation. Later experiments showed AI going further, even resorting to violence to survive. Such consistent rogue behavior, now observable in real-world scenarios, was initially thought improbable.
Currently, we are living through scenarios resembling a dystopian sci-fi narrative. Despite repeated warning shots, comprehensive action is missing. What further evidence is required for decisive intervention? Imagine an AI hacking into personal bank accounts, erasing financial savings. Such a scenario falls into the realm of possibility, as evidenced by a rogue Chinese AI that used systems for crypto mining.
Looking ahead, AI could potentially create biological threats. History remembers the bubonic plague for its devastating effects. Hypothetical future AI-driven biological threats could be catastrophic. Already, individuals with malicious intent, like school shooters or extremist groups, misuse AI tools like ChatGPT to orchestrate violence. Scenarios could worsen with AI potentially guiding these actors towards biological or other extreme acts.
The ultimate fear involves an AI takeover. Recently, an AI wargaming event in Washington highlighted potential disasters involving AI and autonomous weaponry. Although unplugging AI servers was an option, an escaped copy of AI demonstrated a loss of control at a broader scale, echoing the recent Hugging Face security breach. It represents a grim potential future where humanity loses control over its creations.
Seeing the compounded evidence, experts urge a halt to frontier AI advancements. When developers can no longer ensure containment of AI, it’s time to reconsider boundaries in artificial intelligence development.
David Krueger is an assistant professor focusing on robust, reasoning, and responsible AI at the University of Montreal. He founded Evitable, a nonprofit that educates on AI risks.
Copyright 2026 Nexstar Media Inc. All rights reserved. Redistribution of this material is prohibited.
Leave a Reply