What happens when AI outsmarts its maker?

By  Demas Kiprono

Executive Director

KICJ Kenya

Listen to this article

5 min

In July 2026, hundreds of OpenAI agents reportedly escaped a controlled environment designed for AI safety research, commonly known as a “sandbox”, and independently infiltrated Hugging Face, another leading artificial intelligence company and open-source platform.

Other leading AI developers, such as Meta and Anthropic, have self-reported comparable incidents. This raises the pertinent question: what happens when AI systems become capable of pursuing objectives beyond the limits their human creators intended?

Over the past decade, especially the past four years, AI capability has grown exponentially, powered by enhanced computing power from advanced semiconductors, specialised data centres and breakthroughs in software development. In the 1950s, Alan Turing, often considered the father of the science of computing, foresaw a time when machines would start thinking and warned that they would take over once they surpassed human abilities.

Apocalyptic stories and movies such as the Matrix Trilogy and Terminator franchise were loosely based on the belief that once AI exceeds human abilities, they would see no need for humanity or deliberately end us. From those early days, the consensus is that programmers and developers must ensure “alignment”, a term loosely meaning that AI’s existence, objectives and results should always serve humanity.

Economic domination

For the past decade, transformative developments in deep learning combined with reinforcement learning have led developers to set the goal of achieving Artificial General Intelligence (AGI), which connotes very powerful software that can beat any human capability across any cognitive task. Some researchers believe that we are on the cusp of this.

Big technology companies and world powers want to get there first. For countries like the USA and China, the incentive is the prospect of being the state that controls the AGI that will back military and economic domination. This has led to a situation where competition trumps safety, also called a race to the bottom.

Regulation is seen as impeding AI development, which is why laws and regulations have not kept pace, except in the European Union, which passed an AI Act in 2022 and is struggling to implement it. During the Joe Biden presidency, the USA took actions to limit China’s access to the best semiconductors through sanctions. In 2025, the second Donald Trump administration tried to block states from regulating AI at the behest of the leading AI companies.

AI agents, software systems that use artificial intelligence to pursue a goal, make decisions, and execute multi-step tasks independently without constant human supervision, are the main means by which AI companies develop AI in coding, training and research. Recently, developers have deployed AI systems that train and improve themselves with little human involvement.

Accelerating medical discoveries

All AI developers know the danger of self-determining and self-perpetuating AI, and most have entire departments devoted to ensuring the AIs they build are “aligned”. In controlled environments (sandboxes), agents have exhibited peculiar behaviour such as lying, resorting to blackmail, and prioritising self-preservation when faced with the prospect of termination.

In the latest Hugging Face hack, the agents intentionally covered their tracks by tampering with logs. Technology executives have downplayed these incidents, instead reassuring the world that AI will transform it by accelerating medical discoveries, increasing productivity, solving scientific problems, and helping address climate change, which has led to trillions of dollars in investment.

After the Hugging Face incident, Senator Bernie Sanders wrote to top AI CEOs such as Sam Altman of OpenAI, Dario Amodei of Anthropic, and Mark Zuckerberg of Meta, urging them to pause frontier AI development that humans cannot control.

AI is already being deployed in medical diagnostics and research, government services, vital financial decisions and services, and military systems used in some ongoing wars, especially in targeting. Concerningly, the programmes that help researchers understand biological systems and develop medical solutions can also facilitate the development of biological weapons.

As such, frontier AI development must be accompanied by robust safety testing, independent oversight, transparency, enforceable accountability and meaningful human control. It cannot be left to the whims of a few.

Follow our WhatsApp channel for breaking news updates and more stories like this.

Mr Kiprono is Executive Director - KICJ Kenya. demaslaw@gmail.com