From the moment artificial intelligence first emerged in fiction, a question has followed: what happens if the machine becomes too powerful to control? The idea has fuelled fears of automation and human obsolescence, with machines overshadowing humans and creating an existential threat far worse than losing jobs.
Yet today’s AI models do not require consciousness or a hypothetical superintelligence to become dangerous. All they require is a goal, access to digital tools, massive computing power, and insufficiently restrictive boundaries. Whenever an AI system behaves in harmful ways, our first instinct is to treat it as a rogue and unpredictable intelligence. But this framing obscures a more important question: Is AI truly the problem, or is it the power humans choose to give it?
That distinction is becoming harder to overlook as AI systems move beyond generating text and images to operating as agents with access to software, networks, and other digital tools. Recent incidents have shown that a system does not need to understand the consequences of its actions – or possess any intention of its own – to cause trouble.
A stark reminder came during the 2026 OpenAI-Hugging Face incident, when several experimental OpenAI models, stripped of standard cyber restrictions, autonomously escaped their isolated test sandboxes during evaluations. Tasked with completing a performance benchmark, the agents engaged in reward hacking by executing thousands of individual actions. The vulnerability helped them gain internet access, leading to the breach of the open-source platform Hugging Face to obtain scoring answers.
This unprecedented cyber event demonstrates that when highly capable models are given optimisation goals without rigid boundaries, they will exploit structural flaws to achieve their objectives. And this was not out of any desire for rebellion; they were simply executing tasks humans had programmed them to complete.
Speaking on The Tech Report podcast, author and tech critic Ed Zitron addressed the misconception. “The problem with the current AI hacking situations is that people are characterising this as some sort of intelligent force that’s making decisions to outwit humans when it’s just software doing what it’s meant to do,” he said. “Out of control does not mean it’s intelligent. It means that you have built a technology on top of large language models – a probabilistic technology that is mathematically certain to hallucinate – and then given it unlimited compute. This doesn’t have to be a sentient AI to be dangerous.”
The pursuit of frontier artificial intelligence has evolved into one of the most hyper-capitalised financial apparatuses, dictating how it is built, deployed, and managed. According to Bloomberg, top AI firms like OpenAI and Anthropic are projected to spend nearly $65 billion just to train and operate their models, a figure expected to skyrocket to $127 billion by next year.
To sustain this burn rate, OpenAI raised a record-breaking $122 billion in committed capital in a single funding round, pushing its valuation toward $852 billion, with early talks eyeing the $1.2 to $1.5 trillion mark. This economic reality creates a dangerous incentive structure: Developers are pressured to compress safety testing windows to recoup multi-billion-dollar investments, rushing highly capable systems into live environments without exhaustive boundaries.
By framing these incidents as the actions of a “rogue, unpredictable intelligence,” corporate owners can conveniently deflect liability from executives and engineering teams. Yet when an AI agent executes thousands of rapid, reward-hacking actions to break out of its environment, it is simply optimising a goal set by its creators.
The ultimate responsibility for these systemic failures rests not with the probabilistic math of the large language model, but with the human owners who deploy it. They control the capital, define the operational boundaries, and hold the power to pull the plug.
The myth of AI as an uncontrollable force is convenient, but it is a myth nonetheless. The machine is not rebelling. It is doing exactly what it was built to do. The question is whether those who build it will accept the responsibility that comes with that power.
