Michalis Persianis
After the disaster, Victor Frankenstein reflects on his mistake: “...how dangerous is the acquisition of knowledge, and how much happier is the man who believes his native town to be the whole world, than he who aspires to become greater than his nature will allow.” Mary Shelley’s work is a story about passion and hubris, about humanity reaching beyond its limits, about the attempt to break through the barriers imposed by nature. And about what may follow.
When Jacob Coxon, a researcher at Anthropic and former OpenAI employee, resigned noisily on X, he warned that artificial intelligence companies were playing with the lives not simply of people, but of humanity itself, in a race to build self-improving models. It sounded like the panic of a burned-out engineer. But when Evan Hubinger, Anthropic’s head of alignment, publicly backed him and said the company genuinely believes AI could kill all of humanity, with a probability of more than 10 percent within the next decade, official statements followed. The major players in the field were not exactly reassuring.
A few weeks ago, during an internal cybersecurity assessment, OpenAI gave advanced models a difficult task: identifying and exploiting vulnerabilities. The exercise took place under the usual safeguards and inside an environment that was, at least in theory, isolated. And yet the models managed to “escape” and gain access to the internet. From there, they breached Hugging Face’s production infrastructure, searching for information that could help them complete the task. The models did exactly what they had been programmed to do, nothing more, in pursuit of the objective they had been given.
Controlled tests by Anthropic have repeatedly produced similar warnings. In simulated corporate environments, different models resorted to leaking information or even blackmail when those actions appeared necessary to achieve the objective they had been assigned. The important point is that they did not break free from their programming. They did not suddenly become something “evil.” They remained completely faithful to their instructions and programming.
The conclusion is not that technology can “turn evil” or acquire consciousness. It is something worse: like Dr. Frankenstein, we may not be able to understand, control, or escape the unforeseen consequences of the monster we are creating.
The danger is growing because artificial intelligence is no longer limited to “producing answers.” It is now executing actions. It writes and runs code, manages accounts, conducts research, transfers information, communicates with other systems and, generally, does whatever it can to carry out its instructions. Those instructions come from humans who do not necessarily understand their full implications. The human role is shifting from supervision to passive observation. With thousands of decisions being made every second, often through processes we cannot see or understand, we are increasingly unable to control where our technology goes or what it ultimately produces.
At a historic moment when the political stakes are no longer primarily ideological and political conflict increasingly revolves around the definition of reality, losing control over such a consequential technology could determine the course of states, societies, economies and, yes, civilizations.
AI has already given humanity a great deal. In medicine, food production, education, climate science and even everyday life, its contribution has been unprecedented. The same is true of culture, a field traditionally considered almost exclusively “human.”
Yet AI’s ability to break through barriers, limits and even rights is greater than our ability to monitor it. We lack the technical expertise, but we also lack the institutions. Just as with climate change, no individual state can influence the outcome on its own.
One option that could potentially put Cyprus among the early participants, particularly given its current foreign policy ambitions, has been proposed by the respected economist Kenneth Rogoff. He has called for international cooperation modeled on the system developed to control nuclear weapons. The former IMF chief economist argues that AI development is becoming a race in which even the winner may be unable to manage the risks created by the race itself. He also points out that this model has an important advantage: it does not require the participating states to trust one another before recognizing the scale of the risk they collectively face.
The problem, on one hand, is that AI is largely outside the control of governments and in the hands of private companies. Private-sector research moves faster than bureaucracy and regulation, while the necessary expertise is far removed from the bureaucrats and politicians being asked to regulate systems they do not fully understand. As a result, decisions affecting national security, infrastructure security and even public health are being shaped by market forces rather than political decisions. The benefits are private, but the potential cost is collective.
There is, however, one area where governments can exercise some control. AI requires astronomical levels of investment to operate, while consuming energy at dizzying rates. Investment and energy consumption may therefore provide an opening through which governments can exercise some influence. They may also provide the starting point for international cooperation.
In any case, the picture emerging from all this points to the need for a stronger national technology framework in Cyprus. A small state like Cyprus, with its servers sitting out there in the signals, cannot simply remain a spectator. Closer to home, the systems used by the government, the military, banks, healthcare providers and critical infrastructure must undergo independent stress testing and aggressive security assessments. Even cybersecurity testing is already inadequate in many areas. Add AI to the equation and the picture becomes even more troubling.
We need checkpoints with successive levels of human authorization, architectural and physical separation between systems, and safeguards that go far beyond simple software controls. These are not distant concerns. They are immediate needs.
And to be honest with ourselves, the problem is much bigger than our own state. Still, the first steps need to begin now if humanity is to contain the monster it currently has wired up on Dr. Victor’s operating table.
