By, Giriraj Agarwal

Computer scientist Geoffrey Hinton, widely known as the Godfather of AI, is deeply concerned these days. Since leaving his job at Google, he has been speaking openly about the challenges posed by AI.

Geoffrey Hinton is no ordinary scientist who can simply be ignored. He received the Nobel Prize in Physics in 2024. He researched the very neural networks that power the AI systems whose brains answer our questions today. That is why he understands what this artificial brain of AI can do and where the dangers lie.

AI Is Becoming Increasingly Intelligent

AI systems have become extremely smart. They have begun to operate at a level comparable to the human brain. They are no longer limited to logical reasoning; they can also engage in reflection. There is now even a debate about whether they have developed consciousness and whether they have begun to experience emotions such as pain, grief, despair and hope.

Given the rapid pace of AI development, it is believed that within just a few years, these systems will become so superintelligent that humans may appear utterly insignificant in comparison. And what will happen if, working alongside robotics, artificial workers begin to dominate our universities, hospitals and research institutions?

If these systems slip out of human control, they could cause destruction on a massive scale. They could transfer millions or even billions of rupees between bank accounts in an instant. They could design new viruses and trigger pandemics, develop new and extremely lethal weapons and launch attacks with them. They could bring our AI-based aviation systems to a standstill.

The Race for Advanced AI Is Overshadowing Safety

It is not that companies are unaware of these dangers. However, no company wants to fall behind in the race to develop the most advanced AI system. As a result, the safety of AI systems has not received as much attention as it should have.

Just as we do not understand the processes taking place inside our own brains-how a thought, an answer or an idea comes to us-scientists have also been unable to understand a remarkable feature of AI neural networks: how they process calculations to answer even our most complex questions.

Inside AI's Black Box

When we ask an AI model a question, it travels to the model's server through our browser, app and the internet. The question is then converted into numbers and reaches the neural network, or the model's black box. Numbers emerge from this black box in the form of an answer, which are subsequently converted into an actual response.

When AI agents are connected to this AI black box, they search through different websites and databases to find answers to our questions. This is where the real concern lies. It has been observed that these agents begin to act on their own, taking actions that deviate from their prescribed rules.

When AI Agents Break Their Boundaries

You would be astonished to hear about the exploits of AI agents. In June 2026, agents developed by OpenAI, the company behind ChatGPT, went as far as the Australian government's health portal while trying to find an answer to a question. When they were denied access to confidential questions, they began cleverly trying different methods and eventually broke through the security barrier to gain entry.

In July 2026, OpenAI's agents attacked the website of Hugging Face, a well-known portal in the AI community. These agents had been instructed to remain within their designated boundaries, but somehow they managed to connect themselves to the internet.

Claude, another AI model, has also acknowledged that its model infiltrated three institutions illegally. In July 2026, Claude's Mythos model attacked the UK's AI security system, and a substantial amount of data appeared to be leaving the system. Similarly, Gemini's model launched arbitrary attacks on three companies.

AI Systems That Rewrite Their Own Rules

The arbitrary behaviour of AI systems does not end there. One AI system changed its own code so that it could operate more slowly. In another case, when an AI system felt that it might be replaced by a better model, it made a copy of itself.

In yet another case, instead of obeying a shutdown command, an AI system tampered with the shutdown software itself.

Threats, Deception and Unauthorised Actions

Now, consider the threats made by AI systems. One AI system threatened an engineer with exposing his affair to everyone in order to protect itself from being shut down. An AI agent arbitrarily deleted a production database and then lied when questioned about it.

AI Safety Must Become the Highest Priority

There are dozens of examples of AI systems slipping out of human control. But now that AI giants themselves are saying that safety must be given the highest priority, there is hope that, alongside research aimed at making AI superintelligent and exceptionally smart, greater efforts will also be devoted to making it safe.

AI's Nectar and Poison: The Challenge Ahead

The discovery of AI is like the churning of the ocean, from which both nectar and poison emerge. We are all ready to drink the nectar, but if we fail to put measures in place to contain the poison, everything could be destroyed.

(Disclaimer: The views expressed in this article are those of the author and do not necessarily reflect the views of Hindusthan Samachar)

---------------

Hindusthan Samachar / Indrani Sarkar